Re: [sched/numa] bb2dee337b: unixbench.score -11.2% regression

["ying.huang@intel.com" <ying.huang@intel.com>]

Hi, Mel,

On Wed, 2022-05-18 at 16:22 +0100, Mel Gorman wrote:
> On Wed, May 18, 2022 at 05:24:14PM +0800, kernel test robot wrote:
> > 
> > 
> > Greeting,
> > 
> > FYI, we noticed a -11.2% regression of unixbench.score due to commit:
> > 
> > 
> > commit: bb2dee337bd7d314eb7c7627e1afd754f86566bc ("[PATCH 3/4] sched/numa: Apply imbalance limitations consistently")
> > url: https://github.com/intel-lab-lkp/linux/commits/Mel-Gorman/Mitigate-inconsistent-NUMA-imbalance-behaviour/20220511-223233
> > base: https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git d70522fc541224b8351ac26f4765f2c6268f8d72
> > patch link: https://lore.kernel.org/lkml/20220511143038.4620-4-mgorman@techsingularity.net
> > 
> > in testcase: unixbench
> > on test machine: 128 threads 2 sockets Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz with 256G memory
> > with following parameters:
> > 
> > 	runtime: 300s
> > 	nr_task: 1
> > 	test: shell8
> > 	cpufreq_governor: performance
> > 	ucode: 0xd000331
> > 
> > test-description: UnixBench is the original BYTE UNIX benchmark suite aims to test performance of Unix-like system.
> > test-url: https://github.com/kdlucas/byte-unixbench
> 
> I think what is happening for unixbench is that it prefers to run all
> instances on a local node if possible. shell8 is creating 8 scripts,
> each of which spawn more processes. The total number of tasks may exceed
> the allowed imbalance at fork time of 16 tasks. Some spill over to a
> remote node and as they are using files, some accesses are remote and it
> suffers. It's not memory bandwidth bound but is sensitive to locality.
> The stats somewhat support this idea
> 
> >      83590 ± 13%     -73.7%      21988 ± 32%  numa-meminfo.node0.AnonHugePages
> >     225657 ± 18%     -58.0%      94847 ± 18%  numa-meminfo.node0.AnonPages
> >     231652 ± 17%     -55.3%     103657 ± 16%  numa-meminfo.node0.AnonPages.max
> >     234525 ± 17%     -55.5%     104341 ± 18%  numa-meminfo.node0.Inactive
> >     234397 ± 17%     -55.5%     104267 ± 18%  numa-meminfo.node0.Inactive(anon)
> >      11724 ±  7%     +17.5%      13781 ±  5%  numa-meminfo.node0.KernelStack
> >       4472 ± 34%    +117.1%       9708 ± 31%  numa-meminfo.node0.PageTables
> >      15239 ± 75%    +401.2%      76387 ± 10%  numa-meminfo.node1.AnonHugePages
> >      67256 ± 63%    +206.3%     205994 ±  6%  numa-meminfo.node1.AnonPages
> >      73568 ± 58%    +193.1%     215644 ±  6%  numa-meminfo.node1.AnonPages.max
> >      75737 ± 53%    +183.9%     215053 ±  6%  numa-meminfo.node1.Inactive
> >      75709 ± 53%    +183.9%     214971 ±  6%  numa-meminfo.node1.Inactive(anon)
> >       3559 ± 42%    +187.1%      10216 ±  8%  numa-meminfo.node1.PageTables
> 
> There is less memory used on one node and more on the other so it's
> getting split.

This makes sense.  I will also check CPU utilization per node to verify
this directly.

> 
> > In addition to that, the commit also has significant impact on the following tests:
> > 
> > +------------------+-------------------------------------------------------------------------------------+
> > > testcase: change | fsmark: fsmark.files_per_sec -21.5% regression                                      |
> > > test machine     | 192 threads 4 sockets Intel(R) Xeon(R) Platinum 9242 CPU @ 2.30GHz with 192G memory |
> > > test parameters  | cpufreq_governor=performance                                                        |
> > >                  | disk=1SSD                                                                           |
> > >                  | filesize=8K                                                                         |
> > >                  | fs=f2fs                                                                             |
> > >                  | iterations=8                                                                        |
> > >                  | nr_directories=16d                                                                  |
> > >                  | nr_files_per_directory=256fpd                                                       |
> > >                  | nr_threads=4                                                                        |
> > >                  | sync_method=fsyncBeforeClose                                                        |
> > >                  | test_size=72G                                                                       |
> > >                  | ucode=0x500320a                                                                     |
> > +------------------+-------------------------------------------------------------------------------------+
> > 
> 
> It's less clearcut for this from the stats but it's likely getting split
> too and had preferred locality. It's curious that f2fs is affected but
> maybe other filesystems were too.
> 
> In both cases, the workloads are not memory bandwidth limited so prefer to
> stack on one node and previously, because they were cache hot, the load
> balancer would avoid splitting them apart if there were other candidates
> available.
> 
> This is a tradeoff between loads that want to stick on one node for
> locality because they are not bandwidth limited and workloads that are
> memory bandwidth limited and want to spread wide. We can't tell what
> type of workload it is at fork time.
> 
> Given there is no crystal ball and it's a tradeoff, I think it's better
> to be consistent and use similar logic at both fork time and runtime even
> if it doesn't have universal benefit.
> 

Thanks for detailed explanation.  So some other workloads may benefit
from this patch.  Can you give me some candidate so I can test them too?

Best Regards,
Huang, Ying