[Devel] Re: [PATCH v4 00/11] memcg: per cgroup dirty page accounting

Ciju Rajan K ciju at linux.vnet.ibm.com
Tue Nov 2 12:33:15 PDT 2010


Greg Thelen wrote:
> Andrew Morton <akpm at linux-foundation.org> writes:
>
>   
>> On Fri, 29 Oct 2010 00:09:03 -0700
>> Greg Thelen <gthelen at google.com> wrote:
>>
>> This is cool stuff - it's been a long haul.  One day we'll be
>> nearly-finished and someone will write a book telling people how to use
>> it all and lots of people will go "holy crap".  I hope.
>>
>>     
>>> Limiting dirty memory is like fixing the max amount of dirty (hard to reclaim)
>>> page cache used by a cgroup.  So, in case of multiple cgroup writers, they will
>>> not be able to consume more than their designated share of dirty pages and will
>>> be forced to perform write-out if they cross that limit.
>>>
>>> The patches are based on a series proposed by Andrea Righi in Mar 2010.
>>>
>>> Overview:
>>> - Add page_cgroup flags to record when pages are dirty, in writeback, or nfs
>>>   unstable.
>>>
>>> - Extend mem_cgroup to record the total number of pages in each of the 
>>>   interesting dirty states (dirty, writeback, unstable_nfs).  
>>>
>>> - Add dirty parameters similar to the system-wide  /proc/sys/vm/dirty_*
>>>   limits to mem_cgroup.  The mem_cgroup dirty parameters are accessible
>>>   via cgroupfs control files.
>>>       
>> Curious minds will want to know what the default values are set to and
>> how they were determined.
>>     
>
> When a memcg is created, its dirty limits are set to a copy of the
> parent's limits.  If the new cgroup is a top level cgroup, then it
> inherits from the system parameters (/proc/sys/vm/dirty_*).
>
>   
>>> - Consider both system and per-memcg dirty limits in page writeback when
>>>   deciding to queue background writeback or block for foreground writeback.
>>>
>>> Known shortcomings:
>>> - When a cgroup dirty limit is exceeded, then bdi writeback is employed to
>>>   writeback dirty inodes.  Bdi writeback considers inodes from any cgroup, not
>>>   just inodes contributing dirty pages to the cgroup exceeding its limit.  
>>>       
>> yup.  Some broader discussion of the implications of this shortcoming
>> is needed.  I'm not sure where it would be placed, though. 
>> Documentation/ for now, until you write that book.
>>     
>
> Fair enough.  I can add more text to Documentation/ describing the
> behavior and issue in more detail.
>
>   
>>> - When memory.use_hierarchy is set, then dirty limits are disabled.  This is a
>>>   implementation detail.
>>>       
>> So this is unintentional, and forced upon us my the present implementation?
>>     
>
> Yes, this is not ideal.  I chose not to address this particular issue in
> this series to keep the series smaller.
>
>   
>>>  An enhanced implementation is needed to check the
>>>   chain of parents to ensure that no dirty limit is exceeded.
>>>       
>> How important is it that this be fixed?
>>     
>
> I am not sure if there is interest in hierarchical per-memcg dirty
> limits.  So I don't think that this is very important to be fixed
> immediately.  But the fact that it doesn't work is unexpected.  It would
> be nice if it just worked.  I'll look into making it work.
>
>   
>> And how feasible would that fix be?  A linear walk up the hierarchy
>> list?  More than that?
>>     
>
> I think it should be a simple matter of enhancing
> mem_cgroup_dirty_info() to walk up the hierarchy looking for the cgroup
> closest to its dirty limit.  The only tricky part is that there are
> really two limits (foreground/throttling limit, and a background limit)
> that need to be considered when finding the memcg that most deserves
> inspection by balance_dirty_pages().
>
>   
>>> Performance data:
>>> - A page fault microbenchmark workload was used to measure performance, which
>>>   can be called in read or write mode:
>>>         f = open(foo. $cpu)
>>>         truncate(f, 4096)
>>>         alarm(60)
>>>         while (1) {
>>>                 p = mmap(f, 4096)
>>>                 if (write)
>>> 			*p = 1
>>> 		else
>>> 			x = *p
>>>                 munmap(p)
>>>         }
>>>
>>> - The workload was called for several points in the patch series in different
>>>   modes:
>>>   - s_read is a single threaded reader
>>>   - s_write is a single threaded writer
>>>   - p_read is a 16 thread reader, each operating on a different file
>>>   - p_write is a 16 thread writer, each operating on a different file
>>>
>>> - Measurements were collected on a 16 core non-numa system using "perf stat
>>>   --repeat 3".  The -a option was used for parallel (p_*) runs.
>>>
>>> - All numbers are page fault rate (M/sec).  Higher is better.
>>>
>>> - To compare the performance of a kernel without non-memcg compare the first and
>>>   last rows, neither has memcg configured.  The first row does not include any
>>>   of these memcg patches.
>>>
>>> - To compare the performance of using memcg dirty limits, compare the baseline
>>>   (2nd row titled "w/ memcg") with the the code and memcg enabled (2nd to last
>>>   row titled "all patches").
>>>
>>>                            root_cgroup                    child_cgroup
>>>                  s_read s_write p_read p_write   s_read s_write p_read p_write
>>> mmotm w/o memcg   0.428  0.390   0.429  0.388
>>> mmotm w/ memcg    0.411  0.378   0.391  0.362     0.412  0.377   0.385  0.363
>>> all patches       0.384  0.360   0.370  0.348     0.381  0.363   0.368  0.347
>>> all patches       0.431  0.402   0.427  0.395
>>>   w/o memcg
>>>       
>> afaict this benchmark has demonstrated that the changes do not cause an
>> appreciable performance regression in terms of CPU loading, yes?
>>     
>
> Using the mmap() workload, which is a fault heavy workload...
>
> When memcg is not configured, there is no significant performance
> change.  Depending on the workload the performance is between 0%..3%
> faster.  This is likely workload noise.
>
> When memcg is configured, the performance drops between 4% and 8%.  Some
> of this might be noise, but it is expected that memcg faults will get
> slower because there's more code in the fault path.
>
>   
>> Can we come up with any tests which demonstrate the _benefits_ of the
>> feature?
>>     
>
> Here is a test script that shows a situation where memcg dirty limits
> are beneficial.  The script runs two programs: a dirty page background
> antagonist (dd) and an interactive foreground process (tar).  If the
> scripts argument is false, then both processes are run together in the
> root cgroup sharing system-wide dirty memory in classic fashion.  If the
> script is given a true argument, then a cgroup is used to contain dd
> dirty page consumption.
>
> ---[start]---
> #!/bin/bash
> # dirty.sh - dirty limit performance test script
> echo use_cgroup: $1
>
> # start antagonist
> if $1; then    # if using cgroup to contain 'dd'...
>   mkdir /dev/cgroup/A
>   echo 400M > /dev/cgroup/A/memory.dirty_limit_in_bytes
>   (echo $BASHPID > /dev/cgroup/A/tasks; dd if=/dev/zero of=big.file
>   count=10k bs=1M) &
> else
>   dd if=/dev/zero of=big.file count=10k bs=1M &
> fi
>
> sleep 10
>
> time tar -xzf linux-2.6.36.tar.gz
> wait
> $1 && rmdir /dev/cgroup/A
> ---[end]---
>
> dirty.sh false : dd 59.7MB/s stddev 7.442%, tar 12.2s stddev 25.720%
>   # both in root_cgroup
> dirty.sh true  : dd 55.4MB/s stddev 0.958%, tar  3.8s stddev  0.250%
>   # tar in root_cgroup, dd in cgroup
>   
Reviewed-by: Ciju Rajan K <ciju at linux.vnet.ibm.com>

Tested-by: Ciju Rajan K <ciju at linux.vnet.ibm.com>

> The cgroup reserved dirty memory resources for the rest of the system
> processes (tar in this case).  The tar process had faster and more
> predictable performance.  memcg dirty ratios might be useful to serve
> different task classes (interactive vs batch).  A past discussion
> touched on this: http://lkml.org/lkml/2010/5/20/136
>   

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers




More information about the Devel mailing list