Re: [x86/mce] 7bb39313cd: netperf.Throughput_tps -4.5% regression

[Feng Tang <feng.tang@intel.com>]

On Sat, Jan 16, 2021 at 07:34:26AM -0800, Paul E. McKenney wrote:
> On Sat, Jan 16, 2021 at 11:52:51AM +0800, Feng Tang wrote:
> > Hi Boris,
> > 
> > On Tue, Jan 12, 2021 at 03:14:38PM +0100, Borislav Petkov wrote:
> > > On Tue, Jan 12, 2021 at 10:21:09PM +0800, kernel test robot wrote:
> > > > 
> > > > Greeting,
> > > > 
> > > > FYI, we noticed a -4.5% regression of netperf.Throughput_tps due to commit:
> > > > 
> > > > 
> > > > commit: 7bb39313cd6239e7eb95198950a02b4ad2a08316 ("x86/mce: Make mce_timed_out() identify holdout CPUs")
> > > > https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git ras/core
> > > > 
> > > > 
> > > > in testcase: netperf
> > > > on test machine: 192 threads Intel(R) Xeon(R) Platinum 9242 CPU @ 2.30GHz with 192G memory
> > > > with following parameters:
> > > > 
> > > > 	ip: ipv4
> > > > 	runtime: 300s
> > > > 	nr_threads: 16
> > > > 	cluster: cs-localhost
> > > > 	test: TCP_CRR
> > > > 	cpufreq_governor: performance
> > > > 	ucode: 0x5003003
> > > > 
> > > > test-description: Netperf is a benchmark that can be use to measure various aspect of networking performance.
> > > > test-url: http://www.netperf.org/netperf/
> > > 
> > > I'm very very sceptical this thing benchmarks #MC exception handler
> > > performance. Because the code this patch adds gets run only during a MCE
> > > exception.
> > > 
> > > So unless I'm missing something obvious please check your setup.
> > 
> > We've tracked some similar strange kernel performance changes, like
> > another mce related one [1]. For many of them, the root cause is
> > the patch changes the code or data alignment/address of other
> > components, as could be seen from System.map file.
> > 
> > We added debug patch trying to force data sections of each .o be
> > aligned (isolating components), and run the test 3 times, and
> > the regression is gone.
> > 
> >          %stddev     %change         %stddev
> >              \          |                \  
> >     263059            -0.2%     262523        netperf.Throughput_total_tps
> >      16441            -0.2%      16407        netperf.Throughput_tps
> > 
> > So the -4.5% is likely to be caused by data address change. 
> > 
> > But still there is something I don't understand, that the patch
> > introduces a new cpumask 'mce_missing_cpus', which is 1024B, and
> > from the System.map, all data following it get a 1024B offset,
> > without changing the cacheline alignment situation.
> > 
> > 2 original system map files are attached in case people want
> > to check.
> > 
> > [1]. https://lore.kernel.org/lkml/20200425114414.GU26573@shao2-debian/
> 
> One possibility is that the data-address changes put more stress on the
> TLB, for example, if that region of memory is not covered by a huge
> TLB entry.  If this is the case, is there a convenient way to define
> mce_missing_cpus so as to get it out of the way?

Yes! I also tried some experiment for dTLB, by adding 3 more cpumask_t
right after 'mce_missing_cpus', so that the total offset will be 4KB.
I expected the regression could be gone, but it turns out to have 
a +2.4% improvement.
 
     16741            -4.5%      15980            +2.4%      17149        netperf.Throughput_tps

Which is still kind of out of our control :) 

Thanks,
Feng


> 							Thanx, Paul