[RFC PATCH 0/2] Hot page promotion optimization for large address space

[RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

In order to check how efficiently the existing NUMA balancing
based hot page promotion mechanism can detect hot regions and
promote pages for workloads with large memory footprints, I
wrote and tested a program that allocates huge amount of
memory but routinely touches only small parts of it.

This microbenchmark provisions memory both on DRAM node and CXL node.
It then divides the entire allocated memory into chunks of smaller
size and randomly choses a chunk for generating memory accesses.
Each chunk is then accessed for a fixed number of iterations to
create the notion of hotness. Within each chunk, the individual
pages at 4K granularity are again accessed in random fashion.

When a chunk is taken up for access in this manner, its pages
can either be residing on DRAM or CXL. In the latter case, the NUMA
balancing driven hot page promotion logic is expected to detect and
promote the hot pages that reside on CXL.

The experiment was conducted on a 2P AMD Bergamo system that has
CXL as the 3rd node.

$ numactl -H
available: 3 nodes (0-2)
node 0 cpus: 0-127,256-383
node 0 size: 128054 MB
node 1 cpus: 128-255,384-511
node 1 size: 128880 MB
node 2 cpus:
node 2 size: 129024 MB
node distances:
node   0   1   2 
  0:  10  32  60 
  1:  32  10  50 
  2:  255  255  10

It is seen that number of pages that get promoted is really low and
the reason for it happens to be that the NUMA hint fault latency turns
out to be much higher than the hot threshold most of the times. Here
are a few latency and threshold sample values captured from
should_numa_migrate_memory() routine when the benchmark was run:

latency	threshold (in ms)
20620	1125
56185	1125
98710	1250
148871	1375
182891	1625
369415	1875
630745	2000

The NUMA hint fault latency metric, which is based on absolute time
difference between scanning time and fault time may not be suitable
for applications that have large amounts of memory. If the time
difference between the scan time PTE update and the subsequent access
(hint fault) is more, the existing logic in should_numa_migrate_memory()
to determine if the page needs to be migrated, will exclude more
pages than it selects pages for promotion.

To address this problem, this RFC converts the absolute time based
hint fault latency in to a relative metric. The number of hint faults
that have occurred between the scan time and the page's fault time
is used as the threshold.

This is quite an experimental work and there are things to take
care of still. While more testing needs to be conducted with different
benchmarks, I am posting the patchset here to just get early feedback.

Microbenchmark
==============
Total allocation is 192G which initially occupies full of Node 1 (DRAM)
and half of Node 2 (CXL)
Chunk size is 1G

			Default		Patched

Benchmark score (us)	637,787,351	571,350,410 (-10.41%)
(Lesser is better)

numa_pte_updates	29,834,747	29,275,489
numa_hint_faults	12,512,736	12,080,772
numa_hint_faults_local	0		0
numa_pages_migrated	1,804,583	6,709,580
pgpromote_success	1,804,500	6,709,526
pgpromote_candidate	1,916,720	7,523,345
pgdemote_kswapd		5,358,119	9,438,006
pgdemote_direct		0		0

				Default		Patched
Number of times
should_numa_migrate_memory()
was invoked:			12,512,736	12,080,772

Number of times the migration
request was rejected due to
hint fault latency being
higher than threshold: 		10,595,933	4,557,401

Redis-memtier
=============
memtier_benchmark -t 512 -n 25000 --ratio 1:1 -c 20 -x 1 --key-pattern R:R
--hide-histogram --distinct-client-seed -d 20000 --pipeline=1000

			Default		Patched

Ops/sec			51,921.16	52,694.55
Hits/sec		21,908.72	22,235.03
Misses/sec		4051.86		4112.24
Avg. Latency		867.51710	591.27561 (-31.84%)
p50 Latency		876.54300	708.60700 (-19.15%)
p99 Latency		1044.47900	1044.47900
p99.9 Latency		1048.57500	1048.57500
KB/sec			937,330.19	951,291.76

numa_pte_updates	66,628,064	72,125,512
numa_hint_faults	57,093,369	63,369,538
numa_hint_faults_local	0		0
numa_pages_migrated	799,128		3,634,114
pgpromote_success	798,974		3,633,672
pgpromote_candidate	33,884,196	23,143,552
pgdemote_kswapd		13,321,784	11,948,894
pgdemote_direct		257		57,147

Bharata B Rao (2):
  sched/numa: Fault count based NUMA hint fault latency
  mm: Update hint fault count for pages that are skipped during scanning

 include/linux/mm.h       | 23 ++++---------
 include/linux/mm_types.h |  3 ++
 kernel/sched/debug.c     |  2 +-
 kernel/sched/fair.c      | 73 +++++++++++-----------------------------
 kernel/sched/sched.h     |  1 +
 mm/huge_memory.c         | 10 +++---
 mm/memory.c              |  2 ++
 mm/mprotect.c            | 14 ++++----
 8 files changed, 46 insertions(+), 82 deletions(-)

-- 
2.25.1

[RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[Bharata B Rao]

For memory tiering mode, the hint fault latency determines
if the page is considered hot enough to be promoted. This
latency value depends on absolute time and is the difference
between the scanning time and the fault time. The default
value of the threshold used to categorize the page as hot
is 1s.

When the address space is huge, pages may not be accessed
right after they are scanned. Hence the hint fault latency
is found to be, most of the times, way beyond the current
threshold thereby resulting in a very low number of page
promotions.

To address this problem, convert the absolute time based hint
fault latency in to a relative metric based. Use the number
of hint faults that have occurred between the scan time and
the page's fault time as the threshold.

TODO: The existing threshold adjustment logic, which is based
on time based hint fault latency has been removed for now.

Signed-off-by: Bharata B Rao <bharata@amd.com>
---
 include/linux/mm.h       | 23 ++++---------
 include/linux/mm_types.h |  3 ++
 kernel/sched/debug.c     |  2 +-
 kernel/sched/fair.c      | 73 +++++++++++-----------------------------
 kernel/sched/sched.h     |  1 +
 mm/huge_memory.c         |  3 +-
 mm/memory.c              |  2 ++
 mm/mprotect.c            |  5 +--
 8 files changed, 37 insertions(+), 75 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index f5a97dec5169..cb1e79f2920b 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1656,17 +1656,7 @@ static inline int folio_nid(const struct folio *folio)
 }
 
 #ifdef CONFIG_NUMA_BALANCING
-/* page access time bits needs to hold at least 4 seconds */
-#define PAGE_ACCESS_TIME_MIN_BITS	12
-#if LAST_CPUPID_SHIFT < PAGE_ACCESS_TIME_MIN_BITS
-#define PAGE_ACCESS_TIME_BUCKETS				\
-	(PAGE_ACCESS_TIME_MIN_BITS - LAST_CPUPID_SHIFT)
-#else
-#define PAGE_ACCESS_TIME_BUCKETS	0
-#endif
-
-#define PAGE_ACCESS_TIME_MASK				\
-	(LAST_CPUPID_MASK << PAGE_ACCESS_TIME_BUCKETS)
+#define PAGE_FAULT_COUNT_BUCKETS	16
 
 static inline int cpu_pid_to_cpupid(int cpu, int pid)
 {
@@ -1732,13 +1722,12 @@ static inline void page_cpupid_reset_last(struct page *page)
 }
 #endif /* LAST_CPUPID_NOT_IN_PAGE_FLAGS */
 
-static inline int folio_xchg_access_time(struct folio *folio, int time)
+static inline int folio_xchg_fault_count(struct folio *folio, int count)
 {
-	int last_time;
+	int last_count;
 
-	last_time = folio_xchg_last_cpupid(folio,
-					   time >> PAGE_ACCESS_TIME_BUCKETS);
-	return last_time << PAGE_ACCESS_TIME_BUCKETS;
+	last_count = folio_xchg_last_cpupid(folio, count >> PAGE_FAULT_COUNT_BUCKETS);
+	return last_count << PAGE_FAULT_COUNT_BUCKETS;
 }
 
 static inline void vma_set_access_pid_bit(struct vm_area_struct *vma)
@@ -1756,7 +1745,7 @@ static inline int folio_xchg_last_cpupid(struct folio *folio, int cpupid)
 	return folio_nid(folio); /* XXX */
 }
 
-static inline int folio_xchg_access_time(struct folio *folio, int time)
+static inline int folio_xchg_fault_count(struct folio *folio, int time)
 {
 	return 0;
 }
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 8b611e13153e..280043a08f25 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -922,6 +922,9 @@ struct mm_struct {
 
 		/* numa_scan_seq prevents two threads remapping PTEs. */
 		int numa_scan_seq;
+
+		/* Accummulated number of hint faults */
+		atomic_t hint_faults;
 #endif
 		/*
 		 * An operation with batched TLB flushing is going on. Anything
diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c
index 8d5d98a5834d..cd6367cef6cb 100644
--- a/kernel/sched/debug.c
+++ b/kernel/sched/debug.c
@@ -369,7 +369,7 @@ static __init int sched_init_debug(void)
 	debugfs_create_u32("scan_period_min_ms", 0644, numa, &sysctl_numa_balancing_scan_period_min);
 	debugfs_create_u32("scan_period_max_ms", 0644, numa, &sysctl_numa_balancing_scan_period_max);
 	debugfs_create_u32("scan_size_mb", 0644, numa, &sysctl_numa_balancing_scan_size);
-	debugfs_create_u32("hot_threshold_ms", 0644, numa, &sysctl_numa_balancing_hot_threshold);
+	debugfs_create_u32("fault_count_threshold", 0644, numa, &sysctl_numa_balancing_fault_count_threshold);
 #endif
 
 	debugfs_create_file("debug", 0444, debugfs_sched, NULL, &sched_debug_fops);
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6a16129f9a5c..977584683f5f 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1364,8 +1364,11 @@ unsigned int sysctl_numa_balancing_scan_size = 256;
 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
 unsigned int sysctl_numa_balancing_scan_delay = 1000;
 
-/* The page with hint page fault latency < threshold in ms is considered hot */
-unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
+/*
+ * Page is considered hot if the number of hint faults between scan time and
+ * the page's fault time is less than this threshold.
+ */
+unsigned int sysctl_numa_balancing_fault_count_threshold = 1000000;
 
 struct numa_group {
 	refcount_t refcount;
@@ -1750,25 +1753,20 @@ static bool pgdat_free_space_enough(struct pglist_data *pgdat)
 }
 
 /*
- * For memory tiering mode, when page tables are scanned, the scan
- * time will be recorded in struct page in addition to make page
- * PROT_NONE for slow memory page.  So when the page is accessed, in
- * hint page fault handler, the hint page fault latency is calculated
- * via,
+ * For memory tiering mode, when page tables are scanned, the current
+ * hint fault count will be recorded in struct page in addition to
+ * make page PROT_NONE for slow memory page.  So when the page is
+ * accessed, in hint page fault handler, the hint page fault latency is
+ * calculated via,
  *
- *	hint page fault latency = hint page fault time - scan time
+ * hint page fault latency = current hint fault count - fault count at scan time
  *
  * The smaller the hint page fault latency, the higher the possibility
  * for the page to be hot.
  */
-static int numa_hint_fault_latency(struct folio *folio)
+static inline int numa_hint_fault_latency(struct folio *folio, int count)
 {
-	int last_time, time;
-
-	time = jiffies_to_msecs(jiffies);
-	last_time = folio_xchg_access_time(folio, time);
-
-	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
+	return count - folio_xchg_fault_count(folio, count);
 }
 
 /*
@@ -1794,35 +1792,6 @@ static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
 	return false;
 }
 
-#define NUMA_MIGRATION_ADJUST_STEPS	16
-
-static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
-					    unsigned long rate_limit,
-					    unsigned int ref_th)
-{
-	unsigned int now, start, th_period, unit_th, th;
-	unsigned long nr_cand, ref_cand, diff_cand;
-
-	now = jiffies_to_msecs(jiffies);
-	th_period = sysctl_numa_balancing_scan_period_max;
-	start = pgdat->nbp_th_start;
-	if (now - start > th_period &&
-	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
-		ref_cand = rate_limit *
-			sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
-		nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
-		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
-		unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
-		th = pgdat->nbp_threshold ? : ref_th;
-		if (diff_cand > ref_cand * 11 / 10)
-			th = max(th - unit_th, unit_th);
-		else if (diff_cand < ref_cand * 9 / 10)
-			th = min(th + unit_th, ref_th * 2);
-		pgdat->nbp_th_nr_cand = nr_cand;
-		pgdat->nbp_threshold = th;
-	}
-}
-
 bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
 				int src_nid, int dst_cpu)
 {
@@ -1838,7 +1807,7 @@ bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
 	    !node_is_toptier(src_nid)) {
 		struct pglist_data *pgdat;
 		unsigned long rate_limit;
-		unsigned int latency, th, def_th;
+		unsigned int latency;
 
 		pgdat = NODE_DATA(dst_nid);
 		if (pgdat_free_space_enough(pgdat)) {
@@ -1847,16 +1816,13 @@ bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
 			return true;
 		}
 
-		def_th = sysctl_numa_balancing_hot_threshold;
-		rate_limit = sysctl_numa_balancing_promote_rate_limit << \
-			(20 - PAGE_SHIFT);
-		numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
-
-		th = pgdat->nbp_threshold ? : def_th;
-		latency = numa_hint_fault_latency(folio);
-		if (latency >= th)
+		latency = numa_hint_fault_latency(folio,
+						  atomic_read(&p->mm->hint_faults));
+		if (latency >= sysctl_numa_balancing_fault_count_threshold)
 			return false;
 
+		rate_limit = sysctl_numa_balancing_promote_rate_limit << \
+			(20 - PAGE_SHIFT);
 		return !numa_promotion_rate_limit(pgdat, rate_limit,
 						  folio_nr_pages(folio));
 	}
@@ -3444,6 +3410,7 @@ void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
 			mm->numa_scan_seq = 0;
 		}
+		atomic_set(&mm->hint_faults, 0);
 	}
 	p->node_stamp			= 0;
 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index d2242679239e..f975d643fa6a 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -2564,6 +2564,7 @@ extern unsigned int sysctl_numa_balancing_scan_period_min;
 extern unsigned int sysctl_numa_balancing_scan_period_max;
 extern unsigned int sysctl_numa_balancing_scan_size;
 extern unsigned int sysctl_numa_balancing_hot_threshold;
+extern unsigned int sysctl_numa_balancing_fault_count_threshold;
 #endif
 
 #ifdef CONFIG_SCHED_HRTICK
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 94c958f7ebb5..7e62c3c2bbcb 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2101,8 +2101,7 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 
 		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 		    !toptier)
-			folio_xchg_access_time(folio,
-					       jiffies_to_msecs(jiffies));
+			folio_xchg_fault_count(folio, atomic_read(&mm->hint_faults));
 	}
 	/*
 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
diff --git a/mm/memory.c b/mm/memory.c
index 0bfc8b007c01..43a6358e6d31 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4927,6 +4927,8 @@ static vm_fault_t do_numa_page(struct vm_fault *vmf)
 	pte_t pte, old_pte;
 	int flags = 0;
 
+	atomic_inc(&vma->vm_mm->hint_faults);
+
 	/*
 	 * The "pte" at this point cannot be used safely without
 	 * validation through pte_unmap_same(). It's of NUMA type but
diff --git a/mm/mprotect.c b/mm/mprotect.c
index 81991102f785..30118fd492f4 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -159,8 +159,9 @@ static long change_pte_range(struct mmu_gather *tlb,
 					continue;
 				if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 				    !toptier)
-					folio_xchg_access_time(folio,
-						jiffies_to_msecs(jiffies));
+					folio_xchg_fault_count(folio,
+							atomic_read(&vma->vm_mm->hint_faults));
+
 			}
 
 			oldpte = ptep_modify_prot_start(vma, addr, pte);
-- 
2.25.1

Re: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

[snip]

> @@ -1750,25 +1753,20 @@ static bool pgdat_free_space_enough(struct pglist_data *pgdat)
>  }
>  
>  /*
> - * For memory tiering mode, when page tables are scanned, the scan
> - * time will be recorded in struct page in addition to make page
> - * PROT_NONE for slow memory page.  So when the page is accessed, in
> - * hint page fault handler, the hint page fault latency is calculated
> - * via,
> + * For memory tiering mode, when page tables are scanned, the current
> + * hint fault count will be recorded in struct page in addition to
> + * make page PROT_NONE for slow memory page.  So when the page is
> + * accessed, in hint page fault handler, the hint page fault latency is
> + * calculated via,
>   *
> - *	hint page fault latency = hint page fault time - scan time
> + * hint page fault latency = current hint fault count - fault count at scan time
>   *
>   * The smaller the hint page fault latency, the higher the possibility
>   * for the page to be hot.
>   */
> -static int numa_hint_fault_latency(struct folio *folio)
> +static inline int numa_hint_fault_latency(struct folio *folio, int count)
>  {
> -	int last_time, time;
> -
> -	time = jiffies_to_msecs(jiffies);
> -	last_time = folio_xchg_access_time(folio, time);
> -
> -	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
> +	return count - folio_xchg_fault_count(folio, count);
>  }

I found count is task->mm->hint_faults.  That is a process wide
counting.  How do you connect the hotness of a folio with the count of
hint page fault in the process?  How do you compare the hotness of
folios among different processes?

>  /*
> @@ -1794,35 +1792,6 @@ static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
>  	return false;
>  }
>

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[Bharata B Rao]

On 28-Mar-24 7:26 AM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
> [snip]
> 
>> @@ -1750,25 +1753,20 @@ static bool pgdat_free_space_enough(struct pglist_data *pgdat)
>>  }
>>  
>>  /*
>> - * For memory tiering mode, when page tables are scanned, the scan
>> - * time will be recorded in struct page in addition to make page
>> - * PROT_NONE for slow memory page.  So when the page is accessed, in
>> - * hint page fault handler, the hint page fault latency is calculated
>> - * via,
>> + * For memory tiering mode, when page tables are scanned, the current
>> + * hint fault count will be recorded in struct page in addition to
>> + * make page PROT_NONE for slow memory page.  So when the page is
>> + * accessed, in hint page fault handler, the hint page fault latency is
>> + * calculated via,
>>   *
>> - *	hint page fault latency = hint page fault time - scan time
>> + * hint page fault latency = current hint fault count - fault count at scan time
>>   *
>>   * The smaller the hint page fault latency, the higher the possibility
>>   * for the page to be hot.
>>   */
>> -static int numa_hint_fault_latency(struct folio *folio)
>> +static inline int numa_hint_fault_latency(struct folio *folio, int count)
>>  {
>> -	int last_time, time;
>> -
>> -	time = jiffies_to_msecs(jiffies);
>> -	last_time = folio_xchg_access_time(folio, time);
>> -
>> -	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
>> +	return count - folio_xchg_fault_count(folio, count);
>>  }
> 
> I found count is task->mm->hint_faults.  That is a process wide
> counting.  How do you connect the hotness of a folio with the count of
> hint page fault in the process?  How do you compare the hotness of
> folios among different processes?

The global hint fault count that we already maintain could
be used instead of per-task fault. That should take care
of the concern you mention right?

Regards,
Bharata.

Re: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 28-Mar-24 7:26 AM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>> [snip]
>> 
>>> @@ -1750,25 +1753,20 @@ static bool pgdat_free_space_enough(struct pglist_data *pgdat)
>>>  }
>>>  
>>>  /*
>>> - * For memory tiering mode, when page tables are scanned, the scan
>>> - * time will be recorded in struct page in addition to make page
>>> - * PROT_NONE for slow memory page.  So when the page is accessed, in
>>> - * hint page fault handler, the hint page fault latency is calculated
>>> - * via,
>>> + * For memory tiering mode, when page tables are scanned, the current
>>> + * hint fault count will be recorded in struct page in addition to
>>> + * make page PROT_NONE for slow memory page.  So when the page is
>>> + * accessed, in hint page fault handler, the hint page fault latency is
>>> + * calculated via,
>>>   *
>>> - *	hint page fault latency = hint page fault time - scan time
>>> + * hint page fault latency = current hint fault count - fault count at scan time
>>>   *
>>>   * The smaller the hint page fault latency, the higher the possibility
>>>   * for the page to be hot.
>>>   */
>>> -static int numa_hint_fault_latency(struct folio *folio)
>>> +static inline int numa_hint_fault_latency(struct folio *folio, int count)
>>>  {
>>> -	int last_time, time;
>>> -
>>> -	time = jiffies_to_msecs(jiffies);
>>> -	last_time = folio_xchg_access_time(folio, time);
>>> -
>>> -	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
>>> +	return count - folio_xchg_fault_count(folio, count);
>>>  }
>> 
>> I found count is task->mm->hint_faults.  That is a process wide
>> counting.  How do you connect the hotness of a folio with the count of
>> hint page fault in the process?  How do you compare the hotness of
>> folios among different processes?
>
> The global hint fault count that we already maintain could
> be used instead of per-task fault. That should take care
> of the concern you mention right?

I have plotted the total number of hint faults per second before, and it
changes a lot along the time.  So I don't think it is a good
measurement.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[kernel test robot]

Hi Bharata,

[This is a private test report for your RFC patch.]
kernel test robot noticed the following build errors:

[auto build test ERROR on peterz-queue/sched/core]
[also build test ERROR on linus/master v6.9-rc1 next-20240328]
[cannot apply to akpm-mm/mm-everything tip/sched/core]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Bharata-B-Rao/sched-numa-Fault-count-based-NUMA-hint-fault-latency/20240328-000607
base:   https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git sched/core
patch link:    https://lore.kernel.org/r/20240327160237.2355-2-bharata%40amd.com
patch subject: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency
config: m68k-allmodconfig (https://download.01.org/0day-ci/archive/20240330/202403300858.1nHEhscI-lkp@intel.com/config)
compiler: m68k-linux-gcc (GCC) 13.2.0
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20240330/202403300858.1nHEhscI-lkp@intel.com/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202403300858.1nHEhscI-lkp@intel.com/

All errors (new ones prefixed by >>):

   mm/memory.c: In function 'do_numa_page':
>> mm/memory.c:4804:31: error: 'struct mm_struct' has no member named 'hint_faults'
    4804 |         atomic_inc(&vma->vm_mm->hint_faults);
         |                               ^~
--
   mm/mprotect.c: In function 'change_pte_range':
>> mm/mprotect.c:163:80: error: 'struct mm_struct' has no member named 'hint_faults'
     163 |                                                         atomic_read(&vma->vm_mm->hint_faults));
         |                                                                                ^~


vim +4804 mm/memory.c

  4792	
  4793	static vm_fault_t do_numa_page(struct vm_fault *vmf)
  4794	{
  4795		struct vm_area_struct *vma = vmf->vma;
  4796		struct folio *folio = NULL;
  4797		int nid = NUMA_NO_NODE;
  4798		bool writable = false;
  4799		int last_cpupid;
  4800		int target_nid;
  4801		pte_t pte, old_pte;
  4802		int flags = 0;
  4803	
> 4804		atomic_inc(&vma->vm_mm->hint_faults);
  4805	
  4806		/*
  4807		 * The "pte" at this point cannot be used safely without
  4808		 * validation through pte_unmap_same(). It's of NUMA type but
  4809		 * the pfn may be screwed if the read is non atomic.
  4810		 */
  4811		spin_lock(vmf->ptl);
  4812		if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
  4813			pte_unmap_unlock(vmf->pte, vmf->ptl);
  4814			goto out;
  4815		}
  4816	
  4817		/* Get the normal PTE  */
  4818		old_pte = ptep_get(vmf->pte);
  4819		pte = pte_modify(old_pte, vma->vm_page_prot);
  4820	
  4821		/*
  4822		 * Detect now whether the PTE could be writable; this information
  4823		 * is only valid while holding the PT lock.
  4824		 */
  4825		writable = pte_write(pte);
  4826		if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
  4827		    can_change_pte_writable(vma, vmf->address, pte))
  4828			writable = true;
  4829	
  4830		folio = vm_normal_folio(vma, vmf->address, pte);
  4831		if (!folio || folio_is_zone_device(folio))
  4832			goto out_map;
  4833	
  4834		/* TODO: handle PTE-mapped THP */
  4835		if (folio_test_large(folio))
  4836			goto out_map;
  4837	
  4838		/*
  4839		 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
  4840		 * much anyway since they can be in shared cache state. This misses
  4841		 * the case where a mapping is writable but the process never writes
  4842		 * to it but pte_write gets cleared during protection updates and
  4843		 * pte_dirty has unpredictable behaviour between PTE scan updates,
  4844		 * background writeback, dirty balancing and application behaviour.
  4845		 */
  4846		if (!writable)
  4847			flags |= TNF_NO_GROUP;
  4848	
  4849		/*
  4850		 * Flag if the folio is shared between multiple address spaces. This
  4851		 * is later used when determining whether to group tasks together
  4852		 */
  4853		if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
  4854			flags |= TNF_SHARED;
  4855	
  4856		nid = folio_nid(folio);
  4857		/*
  4858		 * For memory tiering mode, cpupid of slow memory page is used
  4859		 * to record page access time.  So use default value.
  4860		 */
  4861		if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
  4862		    !node_is_toptier(nid))
  4863			last_cpupid = (-1 & LAST_CPUPID_MASK);
  4864		else
  4865			last_cpupid = folio_last_cpupid(folio);
  4866		target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags);
  4867		if (target_nid == NUMA_NO_NODE) {
  4868			folio_put(folio);
  4869			goto out_map;
  4870		}
  4871		pte_unmap_unlock(vmf->pte, vmf->ptl);
  4872		writable = false;
  4873	
  4874		/* Migrate to the requested node */
  4875		if (migrate_misplaced_folio(folio, vma, target_nid)) {
  4876			nid = target_nid;
  4877			flags |= TNF_MIGRATED;
  4878		} else {
  4879			flags |= TNF_MIGRATE_FAIL;
  4880			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
  4881						       vmf->address, &vmf->ptl);
  4882			if (unlikely(!vmf->pte))
  4883				goto out;
  4884			if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
  4885				pte_unmap_unlock(vmf->pte, vmf->ptl);
  4886				goto out;
  4887			}
  4888			goto out_map;
  4889		}
  4890	
  4891	out:
  4892		if (nid != NUMA_NO_NODE)
  4893			task_numa_fault(last_cpupid, nid, 1, flags);
  4894		return 0;
  4895	out_map:
  4896		/*
  4897		 * Make it present again, depending on how arch implements
  4898		 * non-accessible ptes, some can allow access by kernel mode.
  4899		 */
  4900		old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
  4901		pte = pte_modify(old_pte, vma->vm_page_prot);
  4902		pte = pte_mkyoung(pte);
  4903		if (writable)
  4904			pte = pte_mkwrite(pte, vma);
  4905		ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
  4906		update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
  4907		pte_unmap_unlock(vmf->pte, vmf->ptl);
  4908		goto out;
  4909	}
  4910	

-- 
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki

Re: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency

[kernel test robot]

Hi Bharata,

[This is a private test report for your RFC patch.]
kernel test robot noticed the following build errors:

[auto build test ERROR on peterz-queue/sched/core]
[also build test ERROR on linus/master v6.9-rc1 next-20240328]
[cannot apply to akpm-mm/mm-everything tip/sched/core]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Bharata-B-Rao/sched-numa-Fault-count-based-NUMA-hint-fault-latency/20240328-000607
base:   https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git sched/core
patch link:    https://lore.kernel.org/r/20240327160237.2355-2-bharata%40amd.com
patch subject: [RFC PATCH 1/2] sched/numa: Fault count based NUMA hint fault latency
config: arm-defconfig (https://download.01.org/0day-ci/archive/20240330/202403301111.fNcnRPWz-lkp@intel.com/config)
compiler: clang version 14.0.6 (https://github.com/llvm/llvm-project.git f28c006a5895fc0e329fe15fead81e37457cb1d1)
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20240330/202403301111.fNcnRPWz-lkp@intel.com/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202403301111.fNcnRPWz-lkp@intel.com/

All errors (new ones prefixed by >>):

>> mm/memory.c:4804:26: error: no member named 'hint_faults' in 'struct mm_struct'
           atomic_inc(&vma->vm_mm->hint_faults);
                       ~~~~~~~~~~  ^
   1 error generated.
--
>> mm/mprotect.c:163:33: error: no member named 'hint_faults' in 'struct mm_struct'
                                                           atomic_read(&vma->vm_mm->hint_faults));
                                                                        ~~~~~~~~~~  ^
   1 error generated.


vim +4804 mm/memory.c

  4792	
  4793	static vm_fault_t do_numa_page(struct vm_fault *vmf)
  4794	{
  4795		struct vm_area_struct *vma = vmf->vma;
  4796		struct folio *folio = NULL;
  4797		int nid = NUMA_NO_NODE;
  4798		bool writable = false;
  4799		int last_cpupid;
  4800		int target_nid;
  4801		pte_t pte, old_pte;
  4802		int flags = 0;
  4803	
> 4804		atomic_inc(&vma->vm_mm->hint_faults);
  4805	
  4806		/*
  4807		 * The "pte" at this point cannot be used safely without
  4808		 * validation through pte_unmap_same(). It's of NUMA type but
  4809		 * the pfn may be screwed if the read is non atomic.
  4810		 */
  4811		spin_lock(vmf->ptl);
  4812		if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
  4813			pte_unmap_unlock(vmf->pte, vmf->ptl);
  4814			goto out;
  4815		}
  4816	
  4817		/* Get the normal PTE  */
  4818		old_pte = ptep_get(vmf->pte);
  4819		pte = pte_modify(old_pte, vma->vm_page_prot);
  4820	
  4821		/*
  4822		 * Detect now whether the PTE could be writable; this information
  4823		 * is only valid while holding the PT lock.
  4824		 */
  4825		writable = pte_write(pte);
  4826		if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
  4827		    can_change_pte_writable(vma, vmf->address, pte))
  4828			writable = true;
  4829	
  4830		folio = vm_normal_folio(vma, vmf->address, pte);
  4831		if (!folio || folio_is_zone_device(folio))
  4832			goto out_map;
  4833	
  4834		/* TODO: handle PTE-mapped THP */
  4835		if (folio_test_large(folio))
  4836			goto out_map;
  4837	
  4838		/*
  4839		 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
  4840		 * much anyway since they can be in shared cache state. This misses
  4841		 * the case where a mapping is writable but the process never writes
  4842		 * to it but pte_write gets cleared during protection updates and
  4843		 * pte_dirty has unpredictable behaviour between PTE scan updates,
  4844		 * background writeback, dirty balancing and application behaviour.
  4845		 */
  4846		if (!writable)
  4847			flags |= TNF_NO_GROUP;
  4848	
  4849		/*
  4850		 * Flag if the folio is shared between multiple address spaces. This
  4851		 * is later used when determining whether to group tasks together
  4852		 */
  4853		if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
  4854			flags |= TNF_SHARED;
  4855	
  4856		nid = folio_nid(folio);
  4857		/*
  4858		 * For memory tiering mode, cpupid of slow memory page is used
  4859		 * to record page access time.  So use default value.
  4860		 */
  4861		if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
  4862		    !node_is_toptier(nid))
  4863			last_cpupid = (-1 & LAST_CPUPID_MASK);
  4864		else
  4865			last_cpupid = folio_last_cpupid(folio);
  4866		target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags);
  4867		if (target_nid == NUMA_NO_NODE) {
  4868			folio_put(folio);
  4869			goto out_map;
  4870		}
  4871		pte_unmap_unlock(vmf->pte, vmf->ptl);
  4872		writable = false;
  4873	
  4874		/* Migrate to the requested node */
  4875		if (migrate_misplaced_folio(folio, vma, target_nid)) {
  4876			nid = target_nid;
  4877			flags |= TNF_MIGRATED;
  4878		} else {
  4879			flags |= TNF_MIGRATE_FAIL;
  4880			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
  4881						       vmf->address, &vmf->ptl);
  4882			if (unlikely(!vmf->pte))
  4883				goto out;
  4884			if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
  4885				pte_unmap_unlock(vmf->pte, vmf->ptl);
  4886				goto out;
  4887			}
  4888			goto out_map;
  4889		}
  4890	
  4891	out:
  4892		if (nid != NUMA_NO_NODE)
  4893			task_numa_fault(last_cpupid, nid, 1, flags);
  4894		return 0;
  4895	out_map:
  4896		/*
  4897		 * Make it present again, depending on how arch implements
  4898		 * non-accessible ptes, some can allow access by kernel mode.
  4899		 */
  4900		old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
  4901		pte = pte_modify(old_pte, vma->vm_page_prot);
  4902		pte = pte_mkyoung(pte);
  4903		if (writable)
  4904			pte = pte_mkwrite(pte, vma);
  4905		ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
  4906		update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
  4907		pte_unmap_unlock(vmf->pte, vmf->ptl);
  4908		goto out;
  4909	}
  4910	

-- 
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki

[RFC PATCH 2/2] mm: Update hint fault count for pages that are skipped during scanning

[Bharata B Rao]

During scanning, PTE updates are skipped for those pages which
are already marked as PROT_NONE. This is required but update the
scan time fault count so that the fault count which is used to
calculate the latency is kept uptodate based on the recent
scanning iteration.

Signed-off-by: Bharata B Rao <bharata@amd.com>
---
 mm/huge_memory.c | 7 ++++---
 mm/mprotect.c    | 9 +++++----
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 7e62c3c2bbcb..24a4f976323e 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2086,9 +2086,6 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 		if (is_huge_zero_pmd(*pmd))
 			goto unlock;
 
-		if (pmd_protnone(*pmd))
-			goto unlock;
-
 		folio = page_folio(pmd_page(*pmd));
 		toptier = node_is_toptier(folio_nid(folio));
 		/*
@@ -2102,6 +2099,10 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 		    !toptier)
 			folio_xchg_fault_count(folio, atomic_read(&mm->hint_faults));
+
+		if (pmd_protnone(*pmd))
+			goto unlock;
+
 	}
 	/*
 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
diff --git a/mm/mprotect.c b/mm/mprotect.c
index 30118fd492f4..cfd3812302be 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -118,10 +118,6 @@ static long change_pte_range(struct mmu_gather *tlb,
 				int nid;
 				bool toptier;
 
-				/* Avoid TLB flush if possible */
-				if (pte_protnone(oldpte))
-					continue;
-
 				folio = vm_normal_folio(vma, addr, oldpte);
 				if (!folio || folio_is_zone_device(folio) ||
 				    folio_test_ksm(folio))
@@ -162,6 +158,11 @@ static long change_pte_range(struct mmu_gather *tlb,
 					folio_xchg_fault_count(folio,
 							atomic_read(&vma->vm_mm->hint_faults));
 
+				/* Avoid TLB flush if possible */
+				if (pte_protnone(oldpte))
+					continue;
+
+
 			}
 
 			oldpte = ptep_modify_prot_start(vma, addr, pte);
-- 
2.25.1

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> In order to check how efficiently the existing NUMA balancing
> based hot page promotion mechanism can detect hot regions and
> promote pages for workloads with large memory footprints, I
> wrote and tested a program that allocates huge amount of
> memory but routinely touches only small parts of it.
>
> This microbenchmark provisions memory both on DRAM node and CXL node.
> It then divides the entire allocated memory into chunks of smaller
> size and randomly choses a chunk for generating memory accesses.
> Each chunk is then accessed for a fixed number of iterations to
> create the notion of hotness. Within each chunk, the individual
> pages at 4K granularity are again accessed in random fashion.
>
> When a chunk is taken up for access in this manner, its pages
> can either be residing on DRAM or CXL. In the latter case, the NUMA
> balancing driven hot page promotion logic is expected to detect and
> promote the hot pages that reside on CXL.
>
> The experiment was conducted on a 2P AMD Bergamo system that has
> CXL as the 3rd node.
>
> $ numactl -H
> available: 3 nodes (0-2)
> node 0 cpus: 0-127,256-383
> node 0 size: 128054 MB
> node 1 cpus: 128-255,384-511
> node 1 size: 128880 MB
> node 2 cpus:
> node 2 size: 129024 MB
> node distances:
> node   0   1   2 
>   0:  10  32  60 
>   1:  32  10  50 
>   2:  255  255  10
>
> It is seen that number of pages that get promoted is really low and
> the reason for it happens to be that the NUMA hint fault latency turns
> out to be much higher than the hot threshold most of the times. Here
> are a few latency and threshold sample values captured from
> should_numa_migrate_memory() routine when the benchmark was run:
>
> latency	threshold (in ms)
> 20620	1125
> 56185	1125
> 98710	1250
> 148871	1375
> 182891	1625
> 369415	1875
> 630745	2000

The access latency of your workload is 20s to 630s, which appears too
long.  Can you try to increase the range of threshold to deal with that?
For example,

echo 100000 > /sys/kernel/debug/sched/numa_balancing/hot_threshold_ms

[snip]

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 28-Mar-24 11:05 AM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
>> In order to check how efficiently the existing NUMA balancing
>> based hot page promotion mechanism can detect hot regions and
>> promote pages for workloads with large memory footprints, I
>> wrote and tested a program that allocates huge amount of
>> memory but routinely touches only small parts of it.
>>
>> This microbenchmark provisions memory both on DRAM node and CXL node.
>> It then divides the entire allocated memory into chunks of smaller
>> size and randomly choses a chunk for generating memory accesses.
>> Each chunk is then accessed for a fixed number of iterations to
>> create the notion of hotness. Within each chunk, the individual
>> pages at 4K granularity are again accessed in random fashion.
>>
>> When a chunk is taken up for access in this manner, its pages
>> can either be residing on DRAM or CXL. In the latter case, the NUMA
>> balancing driven hot page promotion logic is expected to detect and
>> promote the hot pages that reside on CXL.
>>
>> The experiment was conducted on a 2P AMD Bergamo system that has
>> CXL as the 3rd node.
>>
>> $ numactl -H
>> available: 3 nodes (0-2)
>> node 0 cpus: 0-127,256-383
>> node 0 size: 128054 MB
>> node 1 cpus: 128-255,384-511
>> node 1 size: 128880 MB
>> node 2 cpus:
>> node 2 size: 129024 MB
>> node distances:
>> node   0   1   2 
>>   0:  10  32  60 
>>   1:  32  10  50 
>>   2:  255  255  10
>>
>> It is seen that number of pages that get promoted is really low and
>> the reason for it happens to be that the NUMA hint fault latency turns
>> out to be much higher than the hot threshold most of the times. Here
>> are a few latency and threshold sample values captured from
>> should_numa_migrate_memory() routine when the benchmark was run:
>>
>> latency	threshold (in ms)
>> 20620	1125
>> 56185	1125
>> 98710	1250
>> 148871	1375
>> 182891	1625
>> 369415	1875
>> 630745	2000
> 
> The access latency of your workload is 20s to 630s, which appears too
> long.  Can you try to increase the range of threshold to deal with that?
> For example,
> 
> echo 100000 > /sys/kernel/debug/sched/numa_balancing/hot_threshold_ms

That of course should help. But I was exploring alternatives where the
notion of hotness can be de-linked from the absolute scanning time to
the extent possible. For large memory workloads where only parts of memory
get accessed at once, the scanning time can lag from the actual access
time significantly as the data above shows. Wondering if such cases can
be addressed without having to be workload-specific.

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 28-Mar-24 11:05 AM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>>> In order to check how efficiently the existing NUMA balancing
>>> based hot page promotion mechanism can detect hot regions and
>>> promote pages for workloads with large memory footprints, I
>>> wrote and tested a program that allocates huge amount of
>>> memory but routinely touches only small parts of it.
>>>
>>> This microbenchmark provisions memory both on DRAM node and CXL node.
>>> It then divides the entire allocated memory into chunks of smaller
>>> size and randomly choses a chunk for generating memory accesses.
>>> Each chunk is then accessed for a fixed number of iterations to
>>> create the notion of hotness. Within each chunk, the individual
>>> pages at 4K granularity are again accessed in random fashion.
>>>
>>> When a chunk is taken up for access in this manner, its pages
>>> can either be residing on DRAM or CXL. In the latter case, the NUMA
>>> balancing driven hot page promotion logic is expected to detect and
>>> promote the hot pages that reside on CXL.
>>>
>>> The experiment was conducted on a 2P AMD Bergamo system that has
>>> CXL as the 3rd node.
>>>
>>> $ numactl -H
>>> available: 3 nodes (0-2)
>>> node 0 cpus: 0-127,256-383
>>> node 0 size: 128054 MB
>>> node 1 cpus: 128-255,384-511
>>> node 1 size: 128880 MB
>>> node 2 cpus:
>>> node 2 size: 129024 MB
>>> node distances:
>>> node   0   1   2 
>>>   0:  10  32  60 
>>>   1:  32  10  50 
>>>   2:  255  255  10
>>>
>>> It is seen that number of pages that get promoted is really low and
>>> the reason for it happens to be that the NUMA hint fault latency turns
>>> out to be much higher than the hot threshold most of the times. Here
>>> are a few latency and threshold sample values captured from
>>> should_numa_migrate_memory() routine when the benchmark was run:
>>>
>>> latency	threshold (in ms)
>>> 20620	1125
>>> 56185	1125
>>> 98710	1250
>>> 148871	1375
>>> 182891	1625
>>> 369415	1875
>>> 630745	2000
>> 
>> The access latency of your workload is 20s to 630s, which appears too
>> long.  Can you try to increase the range of threshold to deal with that?
>> For example,
>> 
>> echo 100000 > /sys/kernel/debug/sched/numa_balancing/hot_threshold_ms
>
> That of course should help. But I was exploring alternatives where the
> notion of hotness can be de-linked from the absolute scanning time to

In fact, only relative time from scan to hint fault is recorded and
calculated, we have only limited bits.

> the extent possible. For large memory workloads where only parts of memory
> get accessed at once, the scanning time can lag from the actual access
> time significantly as the data above shows. Wondering if such cases can
> be addressed without having to be workload-specific.

Does it really matter to promote the quite cold pages (accessed every
more than 20s)?  And if so, how can we adjust the current algorithm to
cover that?  I think that may be possible via extending the threshold
range.  And I think that we can find some way to extending the range by
default if necessary.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 28-Mar-24 11:33 AM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
>> On 28-Mar-24 11:05 AM, Huang, Ying wrote:
>>> Bharata B Rao <bharata@amd.com> writes:
>>>
>>>> In order to check how efficiently the existing NUMA balancing
>>>> based hot page promotion mechanism can detect hot regions and
>>>> promote pages for workloads with large memory footprints, I
>>>> wrote and tested a program that allocates huge amount of
>>>> memory but routinely touches only small parts of it.
>>>>
>>>> This microbenchmark provisions memory both on DRAM node and CXL node.
>>>> It then divides the entire allocated memory into chunks of smaller
>>>> size and randomly choses a chunk for generating memory accesses.
>>>> Each chunk is then accessed for a fixed number of iterations to
>>>> create the notion of hotness. Within each chunk, the individual
>>>> pages at 4K granularity are again accessed in random fashion.
>>>>
>>>> When a chunk is taken up for access in this manner, its pages
>>>> can either be residing on DRAM or CXL. In the latter case, the NUMA
>>>> balancing driven hot page promotion logic is expected to detect and
>>>> promote the hot pages that reside on CXL.
>>>>
>>>> The experiment was conducted on a 2P AMD Bergamo system that has
>>>> CXL as the 3rd node.
>>>>
>>>> $ numactl -H
>>>> available: 3 nodes (0-2)
>>>> node 0 cpus: 0-127,256-383
>>>> node 0 size: 128054 MB
>>>> node 1 cpus: 128-255,384-511
>>>> node 1 size: 128880 MB
>>>> node 2 cpus:
>>>> node 2 size: 129024 MB
>>>> node distances:
>>>> node   0   1   2 
>>>>   0:  10  32  60 
>>>>   1:  32  10  50 
>>>>   2:  255  255  10
>>>>
>>>> It is seen that number of pages that get promoted is really low and
>>>> the reason for it happens to be that the NUMA hint fault latency turns
>>>> out to be much higher than the hot threshold most of the times. Here
>>>> are a few latency and threshold sample values captured from
>>>> should_numa_migrate_memory() routine when the benchmark was run:
>>>>
>>>> latency	threshold (in ms)
>>>> 20620	1125
>>>> 56185	1125
>>>> 98710	1250
>>>> 148871	1375
>>>> 182891	1625
>>>> 369415	1875
>>>> 630745	2000
>>>
>>> The access latency of your workload is 20s to 630s, which appears too
>>> long.  Can you try to increase the range of threshold to deal with that?
>>> For example,
>>>
>>> echo 100000 > /sys/kernel/debug/sched/numa_balancing/hot_threshold_ms
>>
>> That of course should help. But I was exploring alternatives where the
>> notion of hotness can be de-linked from the absolute scanning time to
> 
> In fact, only relative time from scan to hint fault is recorded and
> calculated, we have only limited bits.
> 
>> the extent possible. For large memory workloads where only parts of memory
>> get accessed at once, the scanning time can lag from the actual access
>> time significantly as the data above shows. Wondering if such cases can
>> be addressed without having to be workload-specific.
> 
> Does it really matter to promote the quite cold pages (accessed every
> more than 20s)?  And if so, how can we adjust the current algorithm to
> cover that?  I think that may be possible via extending the threshold
> range.  And I think that we can find some way to extending the range by
> default if necessary.

I don't think the pages are cold but rather the existing mechanism fails
to categorize them as hot. This is because the pages were scanned way
before the accesses start happening. When repeated accesses are made to
a chunk of memory that has been scanned a while back, none of those
accesses get classified as hot because the scan time is way behind
the current access time. That's the reason we are seeing the value
of latency ranging from 20s to 630s as shown above.

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 28-Mar-24 11:33 AM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>>> On 28-Mar-24 11:05 AM, Huang, Ying wrote:
>>>> Bharata B Rao <bharata@amd.com> writes:
>>>>
>>>>> In order to check how efficiently the existing NUMA balancing
>>>>> based hot page promotion mechanism can detect hot regions and
>>>>> promote pages for workloads with large memory footprints, I
>>>>> wrote and tested a program that allocates huge amount of
>>>>> memory but routinely touches only small parts of it.
>>>>>
>>>>> This microbenchmark provisions memory both on DRAM node and CXL node.
>>>>> It then divides the entire allocated memory into chunks of smaller
>>>>> size and randomly choses a chunk for generating memory accesses.
>>>>> Each chunk is then accessed for a fixed number of iterations to
>>>>> create the notion of hotness. Within each chunk, the individual
>>>>> pages at 4K granularity are again accessed in random fashion.
>>>>>
>>>>> When a chunk is taken up for access in this manner, its pages
>>>>> can either be residing on DRAM or CXL. In the latter case, the NUMA
>>>>> balancing driven hot page promotion logic is expected to detect and
>>>>> promote the hot pages that reside on CXL.
>>>>>
>>>>> The experiment was conducted on a 2P AMD Bergamo system that has
>>>>> CXL as the 3rd node.
>>>>>
>>>>> $ numactl -H
>>>>> available: 3 nodes (0-2)
>>>>> node 0 cpus: 0-127,256-383
>>>>> node 0 size: 128054 MB
>>>>> node 1 cpus: 128-255,384-511
>>>>> node 1 size: 128880 MB
>>>>> node 2 cpus:
>>>>> node 2 size: 129024 MB
>>>>> node distances:
>>>>> node   0   1   2 
>>>>>   0:  10  32  60 
>>>>>   1:  32  10  50 
>>>>>   2:  255  255  10
>>>>>
>>>>> It is seen that number of pages that get promoted is really low and
>>>>> the reason for it happens to be that the NUMA hint fault latency turns
>>>>> out to be much higher than the hot threshold most of the times. Here
>>>>> are a few latency and threshold sample values captured from
>>>>> should_numa_migrate_memory() routine when the benchmark was run:
>>>>>
>>>>> latency	threshold (in ms)
>>>>> 20620	1125
>>>>> 56185	1125
>>>>> 98710	1250
>>>>> 148871	1375
>>>>> 182891	1625
>>>>> 369415	1875
>>>>> 630745	2000
>>>>
>>>> The access latency of your workload is 20s to 630s, which appears too
>>>> long.  Can you try to increase the range of threshold to deal with that?
>>>> For example,
>>>>
>>>> echo 100000 > /sys/kernel/debug/sched/numa_balancing/hot_threshold_ms
>>>
>>> That of course should help. But I was exploring alternatives where the
>>> notion of hotness can be de-linked from the absolute scanning time to
>> 
>> In fact, only relative time from scan to hint fault is recorded and
>> calculated, we have only limited bits.
>> 
>>> the extent possible. For large memory workloads where only parts of memory
>>> get accessed at once, the scanning time can lag from the actual access
>>> time significantly as the data above shows. Wondering if such cases can
>>> be addressed without having to be workload-specific.
>> 
>> Does it really matter to promote the quite cold pages (accessed every
>> more than 20s)?  And if so, how can we adjust the current algorithm to
>> cover that?  I think that may be possible via extending the threshold
>> range.  And I think that we can find some way to extending the range by
>> default if necessary.
>
> I don't think the pages are cold but rather the existing mechanism fails
> to categorize them as hot. This is because the pages were scanned way
> before the accesses start happening. When repeated accesses are made to
> a chunk of memory that has been scanned a while back, none of those
> accesses get classified as hot because the scan time is way behind
> the current access time. That's the reason we are seeing the value
> of latency ranging from 20s to 630s as shown above.

If repeated accesses continue, the page will be identified as hot when
it is scanned next time even if we don't expand the threshold range.  If
the repeated accesses only last very short time, it makes little sense
to identify the pages as hot.  Right?

The bits to record scan time or hint page fault is limited, so it's
possible for it to overflow anyway.  We scan scale time stamp if
necessary (for example, from 1ms to 10ms).  But it's hard to scale fault
counter.  And nobody can guarantee the frequency of hint page fault must
be less 1/ms, if it's 10/ms, it can record even short interval.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 29-Mar-24 6:44 AM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
<snip>
>> I don't think the pages are cold but rather the existing mechanism fails
>> to categorize them as hot. This is because the pages were scanned way
>> before the accesses start happening. When repeated accesses are made to
>> a chunk of memory that has been scanned a while back, none of those
>> accesses get classified as hot because the scan time is way behind
>> the current access time. That's the reason we are seeing the value
>> of latency ranging from 20s to 630s as shown above.
> 
> If repeated accesses continue, the page will be identified as hot when
> it is scanned next time even if we don't expand the threshold range.  If
> the repeated accesses only last very short time, it makes little sense
> to identify the pages as hot.  Right?

The total allocated memory here is 192G and the chunk size is 1G. Each
time one such 1G chunk is taken up randomly for generating memory accesses.
Within that 1G, 262144 random accesses are performed and 262144 such
accesses are repeated for 512 times. I thought that should be enough
to classify that chunk of memory as hot. But as we see, often times
the scan time is lagging the access time by a large value.

Let me instrument the code further to learn more insights (if possible)
about the scanning/fault time behaviors here.

Leaving the fault count based threshold apart, do you think there is
value in updating the scan time for skipped pages/PTEs during every
scan so that the scan time remains current for all the pages?

> 
> The bits to record scan time or hint page fault is limited, so it's
> possible for it to overflow anyway.  We scan scale time stamp if
> necessary (for example, from 1ms to 10ms).  But it's hard to scale fault
> counter.  And nobody can guarantee the frequency of hint page fault must
> be less 1/ms, if it's 10/ms, it can record even short interval.

Yes, with the approach I have taken, the time factor is out of the
equation and the notion of hotness is purely a factor of the number of
faults (or accesses)

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 29-Mar-24 6:44 AM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
> <snip>
>>> I don't think the pages are cold but rather the existing mechanism fails
>>> to categorize them as hot. This is because the pages were scanned way
>>> before the accesses start happening. When repeated accesses are made to
>>> a chunk of memory that has been scanned a while back, none of those
>>> accesses get classified as hot because the scan time is way behind
>>> the current access time. That's the reason we are seeing the value
>>> of latency ranging from 20s to 630s as shown above.
>> 
>> If repeated accesses continue, the page will be identified as hot when
>> it is scanned next time even if we don't expand the threshold range.  If
>> the repeated accesses only last very short time, it makes little sense
>> to identify the pages as hot.  Right?
>
> The total allocated memory here is 192G and the chunk size is 1G. Each
> time one such 1G chunk is taken up randomly for generating memory accesses.
> Within that 1G, 262144 random accesses are performed and 262144 such
> accesses are repeated for 512 times. I thought that should be enough
> to classify that chunk of memory as hot.

IIUC, some pages are accessed in very short time (maybe within 1ms).
This isn't repeated access in a long period.  I think that pages
accessed repeatedly in a long period are good candidates for promoting.
But pages accessed frequently in only very short time aren't.

> But as we see, often times
> the scan time is lagging the access time by a large value.
>
> Let me instrument the code further to learn more insights (if possible)
> about the scanning/fault time behaviors here.
>
> Leaving the fault count based threshold apart, do you think there is
> value in updating the scan time for skipped pages/PTEs during every
> scan so that the scan time remains current for all the pages?

No, I don't think so.  That makes hint page fault latency more
inaccurate.

>> 
>> The bits to record scan time or hint page fault is limited, so it's
>> possible for it to overflow anyway.  We scan scale time stamp if
>> necessary (for example, from 1ms to 10ms).  But it's hard to scale fault
>> counter.  And nobody can guarantee the frequency of hint page fault must
>> be less 1/ms, if it's 10/ms, it can record even short interval.
>
> Yes, with the approach I have taken, the time factor is out of the
> equation and the notion of hotness is purely a factor of the number of
> faults (or accesses)

Sorry, I don't get your idea here.  I think that the fault count may be
worse than time in quite some cases.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 02-Apr-24 7:33 AM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
>> On 29-Mar-24 6:44 AM, Huang, Ying wrote:
>>> Bharata B Rao <bharata@amd.com> writes:
>> <snip>
>>>> I don't think the pages are cold but rather the existing mechanism fails
>>>> to categorize them as hot. This is because the pages were scanned way
>>>> before the accesses start happening. When repeated accesses are made to
>>>> a chunk of memory that has been scanned a while back, none of those
>>>> accesses get classified as hot because the scan time is way behind
>>>> the current access time. That's the reason we are seeing the value
>>>> of latency ranging from 20s to 630s as shown above.
>>>
>>> If repeated accesses continue, the page will be identified as hot when
>>> it is scanned next time even if we don't expand the threshold range.  If
>>> the repeated accesses only last very short time, it makes little sense
>>> to identify the pages as hot.  Right?
>>
>> The total allocated memory here is 192G and the chunk size is 1G. Each
>> time one such 1G chunk is taken up randomly for generating memory accesses.
>> Within that 1G, 262144 random accesses are performed and 262144 such
>> accesses are repeated for 512 times. I thought that should be enough
>> to classify that chunk of memory as hot.
> 
> IIUC, some pages are accessed in very short time (maybe within 1ms).
> This isn't repeated access in a long period.  I think that pages
> accessed repeatedly in a long period are good candidates for promoting.
> But pages accessed frequently in only very short time aren't.

Here are the numbers for the 192nd chunk:

Each iteration of 262144 random accesses takes around ~10ms
512 such iterations are taking ~5s
numa_scan_seq is 16 when this chunk is accessed.
And no page promotions were done from this chunk. All the
time should_numa_migrate_memory() found the NUMA hint fault
latency to be higher than threshold.

Are these time periods considered too short for the pages
to be detected as hot and promoted?

> 
>> But as we see, often times
>> the scan time is lagging the access time by a large value.
>>
>> Let me instrument the code further to learn more insights (if possible)
>> about the scanning/fault time behaviors here.
>>
>> Leaving the fault count based threshold apart, do you think there is
>> value in updating the scan time for skipped pages/PTEs during every
>> scan so that the scan time remains current for all the pages?
> 
> No, I don't think so.  That makes hint page fault latency more
> inaccurate.

For the case that I have shown, depending on a old value of scan
time doesn't work well when pages get accessed after a long time
since scanning. At least with the scheme I show in patch 2/2,
probability of detecting pages as hot increases.

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 02-Apr-24 7:33 AM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>>> On 29-Mar-24 6:44 AM, Huang, Ying wrote:
>>>> Bharata B Rao <bharata@amd.com> writes:
>>> <snip>
>>>>> I don't think the pages are cold but rather the existing mechanism fails
>>>>> to categorize them as hot. This is because the pages were scanned way
>>>>> before the accesses start happening. When repeated accesses are made to
>>>>> a chunk of memory that has been scanned a while back, none of those
>>>>> accesses get classified as hot because the scan time is way behind
>>>>> the current access time. That's the reason we are seeing the value
>>>>> of latency ranging from 20s to 630s as shown above.
>>>>
>>>> If repeated accesses continue, the page will be identified as hot when
>>>> it is scanned next time even if we don't expand the threshold range.  If
>>>> the repeated accesses only last very short time, it makes little sense
>>>> to identify the pages as hot.  Right?
>>>
>>> The total allocated memory here is 192G and the chunk size is 1G. Each
>>> time one such 1G chunk is taken up randomly for generating memory accesses.
>>> Within that 1G, 262144 random accesses are performed and 262144 such
>>> accesses are repeated for 512 times. I thought that should be enough
>>> to classify that chunk of memory as hot.
>> 
>> IIUC, some pages are accessed in very short time (maybe within 1ms).
>> This isn't repeated access in a long period.  I think that pages
>> accessed repeatedly in a long period are good candidates for promoting.
>> But pages accessed frequently in only very short time aren't.
>
> Here are the numbers for the 192nd chunk:
>
> Each iteration of 262144 random accesses takes around ~10ms
> 512 such iterations are taking ~5s
> numa_scan_seq is 16 when this chunk is accessed.
> And no page promotions were done from this chunk. All the
> time should_numa_migrate_memory() found the NUMA hint fault
> latency to be higher than threshold.
>
> Are these time periods considered too short for the pages
> to be detected as hot and promoted?

Yes.  I think so.  This is burst accessing, not repeated accessing.
IIUC, NUMA balancing based promotion only works for repeated accessing
for long time, for example, >100s.

>> 
>>> But as we see, often times
>>> the scan time is lagging the access time by a large value.
>>>
>>> Let me instrument the code further to learn more insights (if possible)
>>> about the scanning/fault time behaviors here.
>>>
>>> Leaving the fault count based threshold apart, do you think there is
>>> value in updating the scan time for skipped pages/PTEs during every
>>> scan so that the scan time remains current for all the pages?
>> 
>> No, I don't think so.  That makes hint page fault latency more
>> inaccurate.
>
> For the case that I have shown, depending on a old value of scan
> time doesn't work well when pages get accessed after a long time
> since scanning. At least with the scheme I show in patch 2/2,
> probability of detecting pages as hot increases.

Yes.  This may help your cases, but it will hurt other cases with
incorrect hint page fault latency.

To resolve your issue, we can increase the max value of the hot
threshold automatically.  We can work on that if you can find a real
workload.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 03-Apr-24 2:10 PM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
>> On 02-Apr-24 7:33 AM, Huang, Ying wrote:
>>> Bharata B Rao <bharata@amd.com> writes:
>>>
>>>> On 29-Mar-24 6:44 AM, Huang, Ying wrote:
>>>>> Bharata B Rao <bharata@amd.com> writes:
>>>> <snip>
>>>>>> I don't think the pages are cold but rather the existing mechanism fails
>>>>>> to categorize them as hot. This is because the pages were scanned way
>>>>>> before the accesses start happening. When repeated accesses are made to
>>>>>> a chunk of memory that has been scanned a while back, none of those
>>>>>> accesses get classified as hot because the scan time is way behind
>>>>>> the current access time. That's the reason we are seeing the value
>>>>>> of latency ranging from 20s to 630s as shown above.
>>>>>
>>>>> If repeated accesses continue, the page will be identified as hot when
>>>>> it is scanned next time even if we don't expand the threshold range.  If
>>>>> the repeated accesses only last very short time, it makes little sense
>>>>> to identify the pages as hot.  Right?
>>>>
>>>> The total allocated memory here is 192G and the chunk size is 1G. Each
>>>> time one such 1G chunk is taken up randomly for generating memory accesses.
>>>> Within that 1G, 262144 random accesses are performed and 262144 such
>>>> accesses are repeated for 512 times. I thought that should be enough
>>>> to classify that chunk of memory as hot.
>>>
>>> IIUC, some pages are accessed in very short time (maybe within 1ms).
>>> This isn't repeated access in a long period.  I think that pages
>>> accessed repeatedly in a long period are good candidates for promoting.
>>> But pages accessed frequently in only very short time aren't.
>>
>> Here are the numbers for the 192nd chunk:
>>
>> Each iteration of 262144 random accesses takes around ~10ms
>> 512 such iterations are taking ~5s
>> numa_scan_seq is 16 when this chunk is accessed.
>> And no page promotions were done from this chunk. All the
>> time should_numa_migrate_memory() found the NUMA hint fault
>> latency to be higher than threshold.
>>
>> Are these time periods considered too short for the pages
>> to be detected as hot and promoted?
> 
> Yes.  I think so.  This is burst accessing, not repeated accessing.
> IIUC, NUMA balancing based promotion only works for repeated accessing
> for long time, for example, >100s.

Hmm... When a page is accessed 512 times over a period of 5s and it is
still not detected as hot. This is understandable if fresh scanning couldn't
be done as the accesses were bursty and hence they couldn't be captured via
NUMA hint faults. But here the access captured via hint fault is being rejected
as not hot because the scanning was done a while back. But I do see the challenge
here since we depend on scanning time to obtain the frequency-of-access metric.

BTW, for the above same scenario with numa_balancing_mode=1, the remote
accesses get detected and migration to source node is tried. It is a different
matter that eventually pages can't be migrated in this specific scenario as
the src node is already full.

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 03-Apr-24 2:10 PM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>>> On 02-Apr-24 7:33 AM, Huang, Ying wrote:
>>>> Bharata B Rao <bharata@amd.com> writes:
>>>>
>>>>> On 29-Mar-24 6:44 AM, Huang, Ying wrote:
>>>>>> Bharata B Rao <bharata@amd.com> writes:
>>>>> <snip>
>>>>>>> I don't think the pages are cold but rather the existing mechanism fails
>>>>>>> to categorize them as hot. This is because the pages were scanned way
>>>>>>> before the accesses start happening. When repeated accesses are made to
>>>>>>> a chunk of memory that has been scanned a while back, none of those
>>>>>>> accesses get classified as hot because the scan time is way behind
>>>>>>> the current access time. That's the reason we are seeing the value
>>>>>>> of latency ranging from 20s to 630s as shown above.
>>>>>>
>>>>>> If repeated accesses continue, the page will be identified as hot when
>>>>>> it is scanned next time even if we don't expand the threshold range.  If
>>>>>> the repeated accesses only last very short time, it makes little sense
>>>>>> to identify the pages as hot.  Right?
>>>>>
>>>>> The total allocated memory here is 192G and the chunk size is 1G. Each
>>>>> time one such 1G chunk is taken up randomly for generating memory accesses.
>>>>> Within that 1G, 262144 random accesses are performed and 262144 such
>>>>> accesses are repeated for 512 times. I thought that should be enough
>>>>> to classify that chunk of memory as hot.
>>>>
>>>> IIUC, some pages are accessed in very short time (maybe within 1ms).
>>>> This isn't repeated access in a long period.  I think that pages
>>>> accessed repeatedly in a long period are good candidates for promoting.
>>>> But pages accessed frequently in only very short time aren't.
>>>
>>> Here are the numbers for the 192nd chunk:
>>>
>>> Each iteration of 262144 random accesses takes around ~10ms
>>> 512 such iterations are taking ~5s
>>> numa_scan_seq is 16 when this chunk is accessed.
>>> And no page promotions were done from this chunk. All the
>>> time should_numa_migrate_memory() found the NUMA hint fault
>>> latency to be higher than threshold.
>>>
>>> Are these time periods considered too short for the pages
>>> to be detected as hot and promoted?
>> 
>> Yes.  I think so.  This is burst accessing, not repeated accessing.
>> IIUC, NUMA balancing based promotion only works for repeated accessing
>> for long time, for example, >100s.
>
> Hmm... When a page is accessed 512 times over a period of 5s and it is
> still not detected as hot. This is understandable if fresh scanning couldn't
> be done as the accesses were bursty and hence they couldn't be captured via
> NUMA hint faults. But here the access captured via hint fault is being rejected
> as not hot because the scanning was done a while back. But I do see the challenge
> here since we depend on scanning time to obtain the frequency-of-access metric.

Consider some pages that will be accessed once every 1 hour, should we
consider it hot or not?  Will your proposed method deal with that
correctly?

> BTW, for the above same scenario with numa_balancing_mode=1, the remote
> accesses get detected and migration to source node is tried. It is a different
> matter that eventually pages can't be migrated in this specific scenario as
> the src node is already full.

--
Best Regards,
Huang, Ying

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Bharata B Rao]

On 12-Apr-24 12:58 PM, Huang, Ying wrote:
> Bharata B Rao <bharata@amd.com> writes:
> 
>> On 03-Apr-24 2:10 PM, Huang, Ying wrote:
>>>> Here are the numbers for the 192nd chunk:
>>>>
>>>> Each iteration of 262144 random accesses takes around ~10ms
>>>> 512 such iterations are taking ~5s
>>>> numa_scan_seq is 16 when this chunk is accessed.
>>>> And no page promotions were done from this chunk. All the
>>>> time should_numa_migrate_memory() found the NUMA hint fault
>>>> latency to be higher than threshold.
>>>>
>>>> Are these time periods considered too short for the pages
>>>> to be detected as hot and promoted?
>>>
>>> Yes.  I think so.  This is burst accessing, not repeated accessing.
>>> IIUC, NUMA balancing based promotion only works for repeated accessing
>>> for long time, for example, >100s.
>>
>> Hmm... When a page is accessed 512 times over a period of 5s and it is
>> still not detected as hot. This is understandable if fresh scanning couldn't
>> be done as the accesses were bursty and hence they couldn't be captured via
>> NUMA hint faults. But here the access captured via hint fault is being rejected
>> as not hot because the scanning was done a while back. But I do see the challenge
>> here since we depend on scanning time to obtain the frequency-of-access metric.
> 
> Consider some pages that will be accessed once every 1 hour, should we
> consider it hot or not?  Will your proposed method deal with that
> correctly?

The proposed method removes the absolute time as a factor for the decision and instead
relies on the number of hint faults that have occurred since that page was scanned last.
As long as there are enough hint faults happening in that 1 hour (which means a lot many
other accesses have been captured in that 1 hour), that page shouldn't be considered as
hot. You did mention earlier about hint fault rate varying a lot and one thing I haven't
tried yet is to vary the fault threshold based on current or historical fault rate.

Regards,
Bharata.

Re: [RFC PATCH 0/2] Hot page promotion optimization for large address space

[Huang, Ying]

Bharata B Rao <bharata@amd.com> writes:

> On 12-Apr-24 12:58 PM, Huang, Ying wrote:
>> Bharata B Rao <bharata@amd.com> writes:
>> 
>>> On 03-Apr-24 2:10 PM, Huang, Ying wrote:
>>>>> Here are the numbers for the 192nd chunk:
>>>>>
>>>>> Each iteration of 262144 random accesses takes around ~10ms
>>>>> 512 such iterations are taking ~5s
>>>>> numa_scan_seq is 16 when this chunk is accessed.
>>>>> And no page promotions were done from this chunk. All the
>>>>> time should_numa_migrate_memory() found the NUMA hint fault
>>>>> latency to be higher than threshold.
>>>>>
>>>>> Are these time periods considered too short for the pages
>>>>> to be detected as hot and promoted?
>>>>
>>>> Yes.  I think so.  This is burst accessing, not repeated accessing.
>>>> IIUC, NUMA balancing based promotion only works for repeated accessing
>>>> for long time, for example, >100s.
>>>
>>> Hmm... When a page is accessed 512 times over a period of 5s and it is
>>> still not detected as hot. This is understandable if fresh scanning couldn't
>>> be done as the accesses were bursty and hence they couldn't be captured via
>>> NUMA hint faults. But here the access captured via hint fault is being rejected
>>> as not hot because the scanning was done a while back. But I do see the challenge
>>> here since we depend on scanning time to obtain the frequency-of-access metric.
>> 
>> Consider some pages that will be accessed once every 1 hour, should we
>> consider it hot or not?  Will your proposed method deal with that
>> correctly?
>
> The proposed method removes the absolute time as a factor for the decision and instead
> relies on the number of hint faults that have occurred since that page was scanned last.
> As long as there are enough hint faults happening in that 1 hour (which means a lot many
> other accesses have been captured in that 1 hour), that page shouldn't be considered as
> hot. You did mention earlier about hint fault rate varying a lot and one thing I haven't
> tried yet is to vary the fault threshold based on current or historical fault rate.

In your original example, if a lot many other accesses between NUMA
balancing page table scanning and 512 page accesses, you cannot identify
the page as hot too, right?

If the NUMA balancing page table scanning period is much longer than 5s,
it's high possible that we cannot distinguish between 1 and 512 page
accesses within 5s with your method and the original method.

Better discuss the behavior with a more detail example, for example,
when the page is scanned, how many pages are accessed, how long between
accesses, etc.

--
Best Regards,
Huang, Ying