[PATCH RESEND V8 0/2] Currently used jhash are slow enough and replace it allow as to make KSM

[PATCH RESEND V8 0/2] Currently used jhash are slow enough and replace it allow as to make KSM

[Timofey Titovets]

From: Timofey Titovets <timofey.titovets@synesis.ru>

About speed (in kernel):
        ksm: crc32c   hash() 12081 MB/s
        ksm: xxh64    hash()  8770 MB/s
        ksm: xxh32    hash()  4529 MB/s
        ksm: jhash2   hash()  1569 MB/s

By sioh Lee tests (copy from other mail):
Test platform: openstack cloud platform (NEWTON version)
Experiment node: openstack based cloud compute node (CPU: xeon E5-2620 v3, memory 64gb)
VM: (2 VCPU, RAM 4GB, DISK 20GB) * 4
Linux kernel: 4.14 (latest version)
KSM setup - sleep_millisecs: 200ms, pages_to_scan: 200

Experiment process
Firstly, we turn off KSM and launch 4 VMs.
Then we turn on the KSM and measure the checksum computation time until full_scans become two.

The experimental results (the experimental value is the average of the measured values)
crc32c_intel: 1084.10ns
crc32c (no hardware acceleration): 7012.51ns
xxhash32: 2227.75ns
xxhash64: 1413.16ns
jhash2: 5128.30ns

In summary, the result shows that crc32c_intel has advantages over all 
of the hash function used in the experiment. (decreased by 84.54% compared to crc32c,
78.86% compared to jhash2, 51.33% xxhash32, 23.28% compared to xxhash64)
the results are similar to those of Timofey.

But,
use only xxhash for now, because for using crc32c,
cryptoapi must be initialized first - that require some
tricky solution to work good in all situations.

So:
  - Fisrt patch implement compile time pickup of fastest implementation of xxhash
    for target platform.
  - Second replace jhash2 with xxhash

Thanks.

CC: Andrea Arcangeli <aarcange@redhat.com>
CC: linux-mm@kvack.org
CC: kvm@vger.kernel.org
CC: leesioh <solee@os.korea.ac.kr>

Timofey Titovets (2):
  xxHash: create arch dependent 32/64-bit xxhash()
  ksm: replace jhash2 with xxhash

 include/linux/xxhash.h | 23 +++++++++++++++++++++++
 mm/Kconfig             |  1 +
 mm/ksm.c               |  4 ++--
 3 files changed, 26 insertions(+), 2 deletions(-)

-- 
2.19.0

[PATCH RESEND V8 1/2] xxHash: create arch dependent 32/64-bit xxhash()

[Timofey Titovets]

xxh32() - fast on both 32/64-bit platforms
xxh64() - fast only on 64-bit platform

Create xxhash() which will pickup fastest version
on compile time.

As result depends on cpu word size,
the main proporse of that - in memory hashing.

Changes:
  v2:
    - Create that patch
  v3 -> v8:
    - Nothing, whole patchset version bump

Signed-off-by: Timofey Titovets <nefelim4ag@gmail.com>
Reviewed-by: Pavel Tatashin <pavel.tatashin@microsoft.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>

CC: Andrea Arcangeli <aarcange@redhat.com>
CC: linux-mm@kvack.org
CC: kvm@vger.kernel.org
CC: leesioh <solee@os.korea.ac.kr>
---
 include/linux/xxhash.h | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/include/linux/xxhash.h b/include/linux/xxhash.h
index 9e1f42cb57e9..52b073fea17f 100644
--- a/include/linux/xxhash.h
+++ b/include/linux/xxhash.h
@@ -107,6 +107,29 @@ uint32_t xxh32(const void *input, size_t length, uint32_t seed);
  */
 uint64_t xxh64(const void *input, size_t length, uint64_t seed);
 
+/**
+ * xxhash() - calculate wordsize hash of the input with a given seed
+ * @input:  The data to hash.
+ * @length: The length of the data to hash.
+ * @seed:   The seed can be used to alter the result predictably.
+ *
+ * If the hash does not need to be comparable between machines with
+ * different word sizes, this function will call whichever of xxh32()
+ * or xxh64() is faster.
+ *
+ * Return:  wordsize hash of the data.
+ */
+
+static inline unsigned long xxhash(const void *input, size_t length,
+				   uint64_t seed)
+{
+#if BITS_PER_LONG == 64
+       return xxh64(input, length, seed);
+#else
+       return xxh32(input, length, seed);
+#endif
+}
+
 /*-****************************
  * Streaming Hash Functions
  *****************************/
-- 
2.19.0

Re: [PATCH RESEND V8 1/2] xxHash: create arch dependent 32/64-bit xxhash()

[Pavel Tatashin]

Hi Andrew,

Can you please accept these patches? They are simple yet provide a good
performance improvement. Timofey has been resending them for a while.

Thank you,
Pasha

On 18-10-23 21:25:53, Timofey Titovets wrote:
> xxh32() - fast on both 32/64-bit platforms
> xxh64() - fast only on 64-bit platform
> 
> Create xxhash() which will pickup fastest version
> on compile time.
> 
> As result depends on cpu word size,
> the main proporse of that - in memory hashing.
> 
> Changes:
>   v2:
>     - Create that patch
>   v3 -> v8:
>     - Nothing, whole patchset version bump
> 
> Signed-off-by: Timofey Titovets <nefelim4ag@gmail.com>
> Reviewed-by: Pavel Tatashin <pavel.tatashin@microsoft.com>
> Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
> 
> CC: Andrea Arcangeli <aarcange@redhat.com>
> CC: linux-mm@kvack.org
> CC: kvm@vger.kernel.org
> CC: leesioh <solee@os.korea.ac.kr>
> ---
>  include/linux/xxhash.h | 23 +++++++++++++++++++++++
>  1 file changed, 23 insertions(+)
> 
> diff --git a/include/linux/xxhash.h b/include/linux/xxhash.h
> index 9e1f42cb57e9..52b073fea17f 100644
> --- a/include/linux/xxhash.h
> +++ b/include/linux/xxhash.h
> @@ -107,6 +107,29 @@ uint32_t xxh32(const void *input, size_t length, uint32_t seed);
>   */
>  uint64_t xxh64(const void *input, size_t length, uint64_t seed);
>  
> +/**
> + * xxhash() - calculate wordsize hash of the input with a given seed
> + * @input:  The data to hash.
> + * @length: The length of the data to hash.
> + * @seed:   The seed can be used to alter the result predictably.
> + *
> + * If the hash does not need to be comparable between machines with
> + * different word sizes, this function will call whichever of xxh32()
> + * or xxh64() is faster.
> + *
> + * Return:  wordsize hash of the data.
> + */
> +
> +static inline unsigned long xxhash(const void *input, size_t length,
> +				   uint64_t seed)
> +{
> +#if BITS_PER_LONG == 64
> +       return xxh64(input, length, seed);
> +#else
> +       return xxh32(input, length, seed);
> +#endif
> +}
> +
>  /*-****************************
>   * Streaming Hash Functions
>   *****************************/
> -- 
> 2.19.0
> 

[PATCH RESEND V8 2/2] ksm: replace jhash2 with xxhash

[Timofey Titovets]

Replace jhash2 with xxhash.

Perf numbers:
Intel(R) Xeon(R) CPU E5-2420 v2 @ 2.20GHz
ksm: crc32c   hash() 12081 MB/s
ksm: xxh64    hash()  8770 MB/s
ksm: xxh32    hash()  4529 MB/s
ksm: jhash2   hash()  1569 MB/s

>From Sioh Lee:
crc32c_intel: 1084.10ns
crc32c (no hardware acceleration): 7012.51ns
xxhash32: 2227.75ns
xxhash64: 1413.16ns
jhash2: 5128.30ns

As jhash2 always will be slower (for data size like PAGE_SIZE).
Don't use it in ksm at all.

Use only xxhash for now, because for using crc32c,
cryptoapi must be initialized first - that require some
tricky solution to work good in all situations.

Thanks.

Changes:
  v1 -> v2:
    - Move xxhash() to xxhash.h/c and separate patches
  v2 -> v3:
    - Move xxhash() xxhash.c -> xxhash.h
    - replace xxhash_t with 'unsigned long'
    - update kerneldoc above xxhash()
  v3 -> v4:
    - Merge xxhash/crc32 patches
    - Replace crc32 with crc32c (crc32 have same as jhash2 speed)
    - Add auto speed test and auto choice of fastest hash function
  v4 -> v5:
    - Pickup missed xxhash patch
    - Update code with compile time choicen xxhash
    - Add more macros to make code more readable
    - As now that only possible use xxhash or crc32c,
      on crc32c allocation error, skip speed test and fallback to xxhash
    - For workaround too early init problem (crc32c not avaliable),
      move zero_checksum init to first call of fastcall()
    - Don't alloc page for hash testing, use arch zero pages for that
  v5 -> v6:
    - Use libcrc32c instead of CRYPTO API, mainly for
      code/Kconfig deps Simplification
    - Add crc32c_available():
      libcrc32c will BUG_ON on crc32c problems,
      so test crc32c avaliable by crc32c_available()
    - Simplify choice_fastest_hash()
    - Simplify fasthash()
    - struct rmap_item && stable_node have sizeof == 64 on x86_64,
      that makes them cache friendly. As we don't suffer from hash collisions,
      change hash type from unsigned long back to u32.
    - Fix kbuild robot warning, make all local functions static
  v6 -> v7:
    - Drop crc32c for now and use only xxhash in ksm.
  v7 -> v8:
    - Remove empty line changes

Signed-off-by: Timofey Titovets <nefelim4ag@gmail.com>
Signed-off-by: leesioh <solee@os.korea.ac.kr>
Reviewed-by: Pavel Tatashin <pavel.tatashin@microsoft.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>

CC: Andrea Arcangeli <aarcange@redhat.com>
CC: linux-mm@kvack.org
CC: kvm@vger.kernel.org
---
 mm/Kconfig | 1 +
 mm/ksm.c   | 4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/mm/Kconfig b/mm/Kconfig
index a550635ea5c3..b5f923081bce 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -297,6 +297,7 @@ config MMU_NOTIFIER
 config KSM
 	bool "Enable KSM for page merging"
 	depends on MMU
+	select XXHASH
 	help
 	  Enable Kernel Samepage Merging: KSM periodically scans those areas
 	  of an application's address space that an app has advised may be
diff --git a/mm/ksm.c b/mm/ksm.c
index 5b0894b45ee5..1a088306ef81 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -25,7 +25,7 @@
 #include <linux/pagemap.h>
 #include <linux/rmap.h>
 #include <linux/spinlock.h>
-#include <linux/jhash.h>
+#include <linux/xxhash.h>
 #include <linux/delay.h>
 #include <linux/kthread.h>
 #include <linux/wait.h>
@@ -1009,7 +1009,7 @@ static u32 calc_checksum(struct page *page)
 {
 	u32 checksum;
 	void *addr = kmap_atomic(page);
-	checksum = jhash2(addr, PAGE_SIZE / 4, 17);
+	checksum = xxhash(addr, PAGE_SIZE, 0);
 	kunmap_atomic(addr);
 	return checksum;
 }
-- 
2.19.0