REPORT: FIO random read performance degradation without "norandommap" option

REPORT: FIO random read performance degradation without "norandommap" option

[Kudryavtsev, Andrey O]

[-- Attachment #1: Type: text/plain, Size: 1033 bytes --]

Colleagues,
I executed 2-hour runs of 4KRR to understand performance changes across the time on the specific very fast NVMe SSD with 1.6TB capacity.
I noticed the side effect of “norandommap” parameter performing full span test on the block device.
Here is the example of the result with random map (I.e. without "norandommap" option) within 120 minutes windows.
[cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
(IOPS in blue)

As soon as I enabled “norandommap” option the curve has changed into the straight line as expected.

Some technical details:
I’m running Centos 7 with 3.18 kernel, SSD of course in the precondition state. FIO 2.2.2 (I unfortunately got higher CPU utilization with 2.2.4 which I’ll report separately).

Config file:
[global]
name=4k random read
filename=/dev/nvme0n1
ioengine=libaio
direct=1
bs=4k
rw=randread
iodepth=16
numjobs=8
buffered=0
size=100%
randrepeat=0
norandommap
refill_buffers
group_reporting
[job1]


--
Andrey Kudryavtsev,
SSD Solution Architect
Intel Corp.


[-- Attachment #2: Screen Shot 2015-01-09 at 2.33.53 PM.png --]
[-- Type: image/png, Size: 15762 bytes --]

RE: REPORT: FIO random read performance degradation without "norandommap" option

[Elliott, Robert (Server Storage)]

> -----Original Message-----
> From: fio-owner@vger.kernel.org [mailto:fio-owner@vger.kernel.org] On
> Behalf Of Kudryavtsev, Andrey O
> Sent: Friday, 09 January, 2015 4:42 PM
> To: fio@vger.kernel.org
> Subject: REPORT: FIO random read performance degradation without
> "norandommap" option
> 
> Colleagues,
> I executed 2-hour runs of 4KRR to understand performance changes across
> the time on the specific very fast NVMe SSD with 1.6TB capacity.
> I noticed the side effect of "norandommap" parameter performing full span
> test on the block device.
> Here is the example of the result with random map (I.e. without
> "norandommap" option) within 120 minutes windows.
> [cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
> (IOPS in blue)
> 
> As soon as I enabled "norandommap" option the curve has changed into the
> straight line as expected.

It takes resources to maintain the random map table.  I always run with
norandommap unless using verify, which has to remember which ones have
been accessed.

Here's a description I gave someone a while back:
With a huge device (e.g., 5.8 TB from RAID-0 made from 16 SSDs),
if you do not use "norandommap", fio allocates a bitmap for all
the disk blocks to keep track of where it has read or written.
It uses this to avoid accessing the same blocks until all
the blocks have been accessed, and to know which blocks it
needs to verify if verify=<something> is enabled.

For 5.8 TB, that is 1562714136 = 1.5 GB.  Not many of those
huge allocations work, so it 
* hangs the system for a while
* generates estimates like
	[eta 1158050440d:06h:50m:22s]
* and eventually reports
	smalloc: failed adding pool
	fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option or set 'softrandommap'. Or give a larger --alloc-size to fio.

fio continues to run after that; I think it verifies only
the devices for which the allocation worked and ignores
the rest.


---
Rob Elliott    HP Server Storage

Re: REPORT: FIO random read performance degradation without "norandommap" option

[Kudryavtsev, Andrey O]

Robert, thanks for the reply.
That�s exactly confirms my findings. Any idea if LBA map is generated for
every job in case of multiple numjobs?

The fact of the performance recovery after degradation was completely
confusing too. I�m glad the solution was so easy.

Jens, are there any chance to add the note into the documentation?

-- 
Andrey Kudryavtsev,

SSD Solution Architect
Intel Corp. 



On 1/9/15, 2:49 PM, "Elliott, Robert (Server Storage)" <Elliott@hp.com>
wrote:

>
>> -----Original Message-----
>> From: fio-owner@vger.kernel.org [mailto:fio-owner@vger.kernel.org] On
>> Behalf Of Kudryavtsev, Andrey O
>> Sent: Friday, 09 January, 2015 4:42 PM
>> To: fio@vger.kernel.org
>> Subject: REPORT: FIO random read performance degradation without
>> "norandommap" option
>>
>> Colleagues,
>> I executed 2-hour runs of 4KRR to understand performance changes across
>> the time on the specific very fast NVMe SSD with 1.6TB capacity.
>> I noticed the side effect of �norandommap� parameter performing full
>>span
>> test on the block device.
>> Here is the example of the result with random map (I.e. without
>> "norandommap" option) within 120 minutes windows.
>> [cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
>> (IOPS in blue)
>>
>> As soon as I enabled �norandommap� option the curve has changed into the
>> straight line as expected.
>
>It takes resources to maintain the random map table.  I always run with
>norandommap unless using verify, which has to remember which ones have
>been accessed.
>
>Here's a description I gave someone a while back:
>With a huge device (e.g., 5.8 TB from RAID-0 made from 16 SSDs),
>if you do not use "norandommap", fio allocates a bitmap for all
>the disk blocks to keep track of where it has read or written.
>It uses this to avoid accessing the same blocks until all
>the blocks have been accessed, and to know which blocks it
>needs to verify if verify=<something> is enabled.
>
>For 5.8 TB, that is 1562714136 = 1.5 GB.  Not many of those
>huge allocations work, so it
>* hangs the system for a while
>* generates estimates like
>        [eta 1158050440d:06h:50m:22s]
>* and eventually reports
>        smalloc: failed adding pool
>        fio: failed allocating random map. If running a large number of
>jobs, try the 'norandommap' option or set 'softrandommap'. Or give a
>larger --alloc-size to fio.
>
>fio continues to run after that; I think it verifies only
>the devices for which the allocation worked and ignores
>the rest.
>
>
>---
>Rob Elliott    HP Server Storage
>
>

Re: REPORT: FIO random read performance degradation without "norandommap" option

[Kulkarni, Vasu]

Excellent!! atleast this explains me how to get out of crazy eta's , I
think the hang may be possible due to low RAM with combination of high
iodepth/blocksize/numjobs.

On Fri, Jan 9, 2015 at 2:49 PM, Elliott, Robert (Server Storage)
<Elliott@hp.com> wrote:
>
>> -----Original Message-----
>> From: fio-owner@vger.kernel.org [mailto:fio-owner@vger.kernel.org] On
>> Behalf Of Kudryavtsev, Andrey O
>> Sent: Friday, 09 January, 2015 4:42 PM
>> To: fio@vger.kernel.org
>> Subject: REPORT: FIO random read performance degradation without
>> "norandommap" option
>>
>> Colleagues,
>> I executed 2-hour runs of 4KRR to understand performance changes across
>> the time on the specific very fast NVMe SSD with 1.6TB capacity.
>> I noticed the side effect of "norandommap" parameter performing full span
>> test on the block device.
>> Here is the example of the result with random map (I.e. without
>> "norandommap" option) within 120 minutes windows.
>> [cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
>> (IOPS in blue)
>>
>> As soon as I enabled "norandommap" option the curve has changed into the
>> straight line as expected.
>
> It takes resources to maintain the random map table.  I always run with
> norandommap unless using verify, which has to remember which ones have
> been accessed.
>
> Here's a description I gave someone a while back:
> With a huge device (e.g., 5.8 TB from RAID-0 made from 16 SSDs),
> if you do not use "norandommap", fio allocates a bitmap for all
> the disk blocks to keep track of where it has read or written.
> It uses this to avoid accessing the same blocks until all
> the blocks have been accessed, and to know which blocks it
> needs to verify if verify=<something> is enabled.
>
> For 5.8 TB, that is 1562714136 = 1.5 GB.  Not many of those
> huge allocations work, so it
> * hangs the system for a while
> * generates estimates like
>         [eta 1158050440d:06h:50m:22s]
> * and eventually reports
>         smalloc: failed adding pool
>         fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option or set 'softrandommap'. Or give a larger --alloc-size to fio.
>
> fio continues to run after that; I think it verifies only
> the devices for which the allocation worked and ignores
> the rest.
>
>
> ---
> Rob Elliott    HP Server Storage
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe fio" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: REPORT: FIO random read performance degradation without "norandommap" option

[Andrey Kuzmin]

On Sat, Jan 10, 2015 at 2:01 AM, Kudryavtsev, Andrey O
<andrey.o.kudryavtsev@intel.com> wrote:
> Robert, thanks for the reply.
> Thatąs exactly confirms my findings. Any idea if LBA map is generated for
> every job in case of multiple numjobs?
>
> The fact of the performance recovery after degradation was completely
> confusing too. Iąm glad the solution was so easy.

My guess is the problem is not not with the option itself, but with
the way you're trying to use it. randommap tracks written blocks to
avoid rewrite, in particular when one wants to verify data integrity
after the run. In read/write workloads, the option might be utilized
to avoid reading unallocated space on a thin-provisioned volume. In
read-only workloads, there is no point in using random map.

Regarding random map RAM footprint, it's quite modest (around 1.01
bits per minimum block size of the workload, 4K in your case), so one
should only see the memory pressure with capacities in the tenths of
TB range.

Regards,
Andrey

>
> Jens, are there any chance to add the note into the documentation?
>
> --
> Andrey Kudryavtsev,
>
> SSD Solution Architect
> Intel Corp.
>
>
>
> On 1/9/15, 2:49 PM, "Elliott, Robert (Server Storage)" <Elliott@hp.com>
> wrote:
>
>>
>>> -----Original Message-----
>>> From: fio-owner@vger.kernel.org [mailto:fio-owner@vger.kernel.org] On
>>> Behalf Of Kudryavtsev, Andrey O
>>> Sent: Friday, 09 January, 2015 4:42 PM
>>> To: fio@vger.kernel.org
>>> Subject: REPORT: FIO random read performance degradation without
>>> "norandommap" option
>>>
>>> Colleagues,
>>> I executed 2-hour runs of 4KRR to understand performance changes across
>>> the time on the specific very fast NVMe SSD with 1.6TB capacity.
>>> I noticed the side effect of łnorandommap˛ parameter performing full
>>>span
>>> test on the block device.
>>> Here is the example of the result with random map (I.e. without
>>> "norandommap" option) within 120 minutes windows.
>>> [cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
>>> (IOPS in blue)
>>>
>>> As soon as I enabled łnorandommap˛ option the curve has changed into the
>>> straight line as expected.
>>
>>It takes resources to maintain the random map table.  I always run with
>>norandommap unless using verify, which has to remember which ones have
>>been accessed.
>>
>>Here's a description I gave someone a while back:
>>With a huge device (e.g., 5.8 TB from RAID-0 made from 16 SSDs),
>>if you do not use "norandommap", fio allocates a bitmap for all
>>the disk blocks to keep track of where it has read or written.
>>It uses this to avoid accessing the same blocks until all
>>the blocks have been accessed, and to know which blocks it
>>needs to verify if verify=<something> is enabled.
>>
>>For 5.8 TB, that is 1562714136 = 1.5 GB.  Not many of those
>>huge allocations work, so it
>>* hangs the system for a while
>>* generates estimates like
>>        [eta 1158050440d:06h:50m:22s]
>>* and eventually reports
>>        smalloc: failed adding pool
>>        fio: failed allocating random map. If running a large number of
>>jobs, try the 'norandommap' option or set 'softrandommap'. Or give a
>>larger --alloc-size to fio.
>>
>>fio continues to run after that; I think it verifies only
>>the devices for which the allocation worked and ignores
>>the rest.
>>
>>
>>---
>>Rob Elliott    HP Server Storage
>>
>>
>
> --
> To unsubscribe from this list: send the line "unsubscribe fio" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: REPORT: FIO random read performance degradation without "norandommap" option

[Kudryavtsev, Andrey O]

Hi Andrey, 
Thank you, it explains the purpose of this feature. Unfortunately it was
not clear to me from the manual, I faulty supposed it increases the
randomness of the random generator distribution. Thanks again!

I’d propose to have it auto set to "norandommap" if “randread” workload is
specified. Does it make sense?

-- 
Andrey Kudryavtsev,

SSD Solution Architect
Intel Corp. 




On 1/10/15, 1:20 AM, "Andrey Kuzmin" <andrey.v.kuzmin@gmail.com> wrote:

>On Sat, Jan 10, 2015 at 2:01 AM, Kudryavtsev, Andrey O
><andrey.o.kudryavtsev@intel.com> wrote:
>> Robert, thanks for the reply.
>> Thatąs exactly confirms my findings. Any idea if LBA map is generated
>>for
>> every job in case of multiple numjobs?
>>
>> The fact of the performance recovery after degradation was completely
>> confusing too. Iąm glad the solution was so easy.
>
>My guess is the problem is not not with the option itself, but with
>the way you're trying to use it. randommap tracks written blocks to
>avoid rewrite, in particular when one wants to verify data integrity
>after the run. In read/write workloads, the option might be utilized
>to avoid reading unallocated space on a thin-provisioned volume. In
>read-only workloads, there is no point in using random map.
>
>Regarding random map RAM footprint, it's quite modest (around 1.01
>bits per minimum block size of the workload, 4K in your case), so one
>should only see the memory pressure with capacities in the tenths of
>TB range.
>
>Regards,
>Andrey
>
>>
>> Jens, are there any chance to add the note into the documentation?
>>
>> --
>> Andrey Kudryavtsev,
>>
>> SSD Solution Architect
>> Intel Corp.
>>
>>
>>
>> On 1/9/15, 2:49 PM, "Elliott, Robert (Server Storage)" <Elliott@hp.com>
>> wrote:
>>
>>>
>>>> -----Original Message-----
>>>> From: fio-owner@vger.kernel.org [mailto:fio-owner@vger.kernel.org] On
>>>> Behalf Of Kudryavtsev, Andrey O
>>>> Sent: Friday, 09 January, 2015 4:42 PM
>>>> To: fio@vger.kernel.org
>>>> Subject: REPORT: FIO random read performance degradation without
>>>> "norandommap" option
>>>>
>>>> Colleagues,
>>>> I executed 2-hour runs of 4KRR to understand performance changes
>>>>across
>>>> the time on the specific very fast NVMe SSD with 1.6TB capacity.
>>>> I noticed the side effect of łnorandommap˛ parameter performing full
>>>>span
>>>> test on the block device.
>>>> Here is the example of the result with random map (I.e. without
>>>> "norandommap" option) within 120 minutes windows.
>>>> [cid:E6872B64-35D1-4447-A0CF-32E6411D9BDB]
>>>> (IOPS in blue)
>>>>
>>>> As soon as I enabled łnorandommap˛ option the curve has changed into
>>>>the
>>>> straight line as expected.
>>>
>>>It takes resources to maintain the random map table.  I always run with
>>>norandommap unless using verify, which has to remember which ones have
>>>been accessed.
>>>
>>>Here's a description I gave someone a while back:
>>>With a huge device (e.g., 5.8 TB from RAID-0 made from 16 SSDs),
>>>if you do not use "norandommap", fio allocates a bitmap for all
>>>the disk blocks to keep track of where it has read or written.
>>>It uses this to avoid accessing the same blocks until all
>>>the blocks have been accessed, and to know which blocks it
>>>needs to verify if verify=<something> is enabled.
>>>
>>>For 5.8 TB, that is 1562714136 = 1.5 GB.  Not many of those
>>>huge allocations work, so it
>>>* hangs the system for a while
>>>* generates estimates like
>>>        [eta 1158050440d:06h:50m:22s]
>>>* and eventually reports
>>>        smalloc: failed adding pool
>>>        fio: failed allocating random map. If running a large number of
>>>jobs, try the 'norandommap' option or set 'softrandommap'. Or give a
>>>larger --alloc-size to fio.
>>>
>>>fio continues to run after that; I think it verifies only
>>>the devices for which the allocation worked and ignores
>>>the rest.
>>>
>>>
>>>---
>>>Rob Elliott    HP Server Storage
>>>
>>>
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe fio" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>--
>To unsubscribe from this list: send the line "unsubscribe fio" in
>the body of a message to majordomo@vger.kernel.org
>More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: REPORT: FIO random read performance degradation without "norandommap" option

[Elliott, Robert (Server Storage)]

> -----Original Message-----
> From: Kudryavtsev, Andrey O [mailto:andrey.o.kudryavtsev@intel.com]
> Sent: Monday, 12 January, 2015 3:06 PM
> To: Andrey Kuzmin
> Cc: Elliott, Robert (Server Storage); fio@vger.kernel.org
> Subject: Re: REPORT: FIO random read performance degradation without
> "norandommap" option
> 
> Hi Andrey,
> Thank you, it explains the purpose of this feature. Unfortunately it was
> not clear to me from the manual, I faulty supposed it increases the
> randomness of the random generator distribution. Thanks again!
> 
> I’d propose to have it auto set to "norandommap" if “randread” workload is
> specified. Does it make sense?
> 

I don't think I'd change it.  For randread, randommap ensures that each
LBA has been accessed once before it repeats reading any of the LBAs. That
could be useful in some cases, like scanning the media (ensuring you've
read each LBA).

---
Rob Elliott    HP Server Storage