There was a large performance regression that was bisected to
commit 611ae8e3 ("x86/tlb: enable tlb flush range support for
x86"). This patch simply changes the default balance point
between a local and global flush for IvyBridge.
In the interest of allowing the tests to be reproduced, this
patch was tested using mmtests 0.15 with the following
configurations
configs/config-global-dhp__tlbflush-performance
configs/config-global-dhp__scheduler-performance
configs/config-global-dhp__network-performance
Results are from two machines
Ivybridge 4 threads: Intel(R) Core(TM) i3-3240 CPU @ 3.40GHz
Ivybridge 8 threads: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz
Page fault microbenchmark showed nothing interesting.
Ebizzy was configured to run multiple iterations and threads.
Thread counts ranged from 1 to NR_CPUS*2. For each thread count,
it ran 100 iterations and each iteration lasted 10 seconds.
Ivybridge 4 threads
3.13.0-rc7 3.13.0-rc7
vanilla altshift-v3
Mean 1 6395.44 ( 0.00%) 6789.09 ( 6.16%)
Mean 2 7012.85 ( 0.00%) 8052.16 ( 14.82%)
Mean 3 6403.04 ( 0.00%) 6973.74 ( 8.91%)
Mean 4 6135.32 ( 0.00%) 6582.33 ( 7.29%)
Mean 5 6095.69 ( 0.00%) 6526.68 ( 7.07%)
Mean 6 6114.33 ( 0.00%) 6416.64 ( 4.94%)
Mean 7 6085.10 ( 0.00%) 6448.51 ( 5.97%)
Mean 8 6120.62 ( 0.00%) 6462.97 ( 5.59%)
Ivybridge 8 threads
3.13.0-rc7 3.13.0-rc7
vanilla altshift-v3
Mean 1 7336.65 ( 0.00%) 7787.02 ( 6.14%)
Mean 2 8218.41 ( 0.00%) 9484.13 ( 15.40%)
Mean 3 7973.62 ( 0.00%) 8922.01 ( 11.89%)
Mean 4 7798.33 ( 0.00%) 8567.03 ( 9.86%)
Mean 5 7158.72 ( 0.00%) 8214.23 ( 14.74%)
Mean 6 6852.27 ( 0.00%) 7952.45 ( 16.06%)
Mean 7 6774.65 ( 0.00%) 7536.35 ( 11.24%)
Mean 8 6510.50 ( 0.00%) 6894.05 ( 5.89%)
Mean 12 6182.90 ( 0.00%) 6661.29 ( 7.74%)
Mean 16 6100.09 ( 0.00%) 6608.69 ( 8.34%)
Ebizzy hits the worst case scenario for TLB range flushing every
time and it shows for these Ivybridge CPUs at least that the
default choice is a poor on. The patch addresses the problem.
Next was a tlbflush microbenchmark written by Alex Shi at
http://marc.info/?l=linux-kernel&m=133727348217113 . It
measures access costs while the TLB is being flushed. The
expectation is that if there are always full TLB flushes that
the benchmark would suffer and it benefits from range flushing
There are 320 iterations of the test per thread count. The
number of entries is randomly selected with a min of 1 and max
of 512. To ensure a reasonably even spread of entries, the full
range is broken up into 8 sections and a random number selected
within that section.
iteration 1, random number between 0-64
iteration 2, random number between 64-128 etc
This is still a very weak methodology. When you do not know
what are typical ranges, random is a reasonable choice but it
can be easily argued that the opimisation was for smaller ranges
and an even spread is not representative of any workload that
matters. To improve this, we'd need to know the probability
distribution of TLB flush range sizes for a set of workloads
that are considered "common", build a synthetic trace and feed
that into this benchmark. Even that is not perfect because it
would not account for the time between flushes but there are
limits of what can be reasonably done and still be doing
something useful. If a representative synthetic trace is
provided then this benchmark could be revisited and the shift values retuned.
Ivybridge 4 threads
3.13.0-rc7 3.13.0-rc7
vanilla altshift-v3
Mean 1 10.50 ( 0.00%) 10.50 ( 0.03%)
Mean 2 17.59 ( 0.00%) 17.18 ( 2.34%)
Mean 3 22.98 ( 0.00%) 21.74 ( 5.41%)
Mean 5 47.13 ( 0.00%) 46.23 ( 1.92%)
Mean 8 43.30 ( 0.00%) 42.56 ( 1.72%)
Ivybridge 8 threads
3.13.0-rc7 3.13.0-rc7
vanilla altshift-v3
Mean 1 9.45 ( 0.00%) 9.36 ( 0.93%)
Mean 2 9.37 ( 0.00%) 9.70 ( -3.54%)
Mean 3 9.36 ( 0.00%) 9.29 ( 0.70%)
Mean 5 14.49 ( 0.00%) 15.04 ( -3.75%)
Mean 8 41.08 ( 0.00%) 38.73 ( 5.71%)
Mean 13 32.04 ( 0.00%) 31.24 ( 2.49%)
Mean 16 40.05 ( 0.00%) 39.04 ( 2.51%)
For both CPUs, average access time is reduced which is good as
this is the benchmark that was used to tune the shift values in
the first place albeit it is now known *how* the benchmark was
used.
The scheduler benchmarks were somewhat inconclusive. They
showed gains and losses and makes me reconsider how stable those
benchmarks really are or if something else might be interfering
with the test results recently.
Network benchmarks were inconclusive. Almost all results were
flat except for netperf-udp tests on the 4 thread machine.
These results were unstable and showed large variations between
reboots. It is unknown if this is a recent problems but I've
noticed before that netperf-udp results tend to vary.
Based on these results, changing the default for Ivybridge seems
like a logical choice.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Tested-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Alex Shi <alex.shi@linaro.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-cqnadffh1tiqrshthRj3Esge@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
f98b7a772a