Re: [PATCH 7/7] x86: mm: set TLB flush tunable to sane value (33)

From: David Nellans
Date: Wed Jul 02 2014 - 14:26:10 EST



On 07/01/2014 11:48 AM, Dave Hansen wrote:
From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>

This has been run through Intel's LKP tests across a wide range
of modern sytems and workloads and it wasn't shown to make a
measurable performance difference positive or negative.

Now that we have some shiny new tracepoints, we can actually
figure out what the heck is going on.

During a kernel compile, 60% of the flush_tlb_mm_range() calls
are for a single page. It breaks down like this:

size percent percent<=
V V V
GLOBAL: 2.20% 2.20% avg cycles: 2283
1: 56.92% 59.12% avg cycles: 1276
2: 13.78% 72.90% avg cycles: 1505
3: 8.26% 81.16% avg cycles: 1880
4: 7.41% 88.58% avg cycles: 2447
5: 1.73% 90.31% avg cycles: 2358
6: 1.32% 91.63% avg cycles: 2563
7: 1.14% 92.77% avg cycles: 2862
8: 0.62% 93.39% avg cycles: 3542
9: 0.08% 93.47% avg cycles: 3289
10: 0.43% 93.90% avg cycles: 3570
11: 0.20% 94.10% avg cycles: 3767
12: 0.08% 94.18% avg cycles: 3996
13: 0.03% 94.20% avg cycles: 4077
14: 0.02% 94.23% avg cycles: 4836
15: 0.04% 94.26% avg cycles: 5699
16: 0.06% 94.32% avg cycles: 5041
17: 0.57% 94.89% avg cycles: 5473
18: 0.02% 94.91% avg cycles: 5396
19: 0.03% 94.95% avg cycles: 5296
20: 0.02% 94.96% avg cycles: 6749
21: 0.18% 95.14% avg cycles: 6225
22: 0.01% 95.15% avg cycles: 6393
23: 0.01% 95.16% avg cycles: 6861
24: 0.12% 95.28% avg cycles: 6912
25: 0.05% 95.32% avg cycles: 7190
26: 0.01% 95.33% avg cycles: 7793
27: 0.01% 95.34% avg cycles: 7833
28: 0.01% 95.35% avg cycles: 8253
29: 0.08% 95.42% avg cycles: 8024
30: 0.03% 95.45% avg cycles: 9670
31: 0.01% 95.46% avg cycles: 8949
32: 0.01% 95.46% avg cycles: 9350
33: 3.11% 98.57% avg cycles: 8534
34: 0.02% 98.60% avg cycles: 10977
35: 0.02% 98.62% avg cycles: 11400

We get in to dimishing returns pretty quickly. On pre-IvyBridge
CPUs, we used to set the limit at 8 pages, and it was set at 128
on IvyBrige. That 128 number looks pretty silly considering that
less than 0.5% of the flushes are that large.

The previous code tried to size this number based on the size of
the TLB. Good idea, but it's error-prone, needs maintenance
(which it didn't get up to now), and probably would not matter in
practice much.

Settting it to 33 means that we cover the mallopt
M_TRIM_THRESHOLD, which is the most universally common size to do
flushes.

That's the short version. Here's the long one for why I chose 33:

1. These numbers have a constant bias in the timestamps from the
tracing. Probably counts for a couple hundred cycles in each of
these tests, but it should be fairly _even_ across all of them.
The smallest delta between the tracepoints I have ever seen is
335 cycles. This is one reason the cycles/page cost goes down in
general as the flushes get larger. The true cost is nearer to
100 cycles.
2. A full flush is more expensive than a single invlpg, but not
by much (single percentages).
3. A dtlb miss is 17.1ns (~45 cycles) and a itlb miss is 13.0ns
(~34 cycles). At those rates, refilling the 512-entry dTLB takes
22,000 cycles.
4. 22,000 cycles is approximately the equivalent of doing 85
invlpg operations. But, the odds are that the TLB can
actually be filled up faster than that because TLB misses that
are close in time also tend to leverage the same caches.
6. ~98% of flushes are <=33 pages. There are a lot of flushes of
33 pages, probably because libc's M_TRIM_THRESHOLD is set to
128k (32 pages)
7. I've found no consistent data to support changing the IvyBridge
vs. SandyBridge tunable by a factor of 16

I used the performance counters on this hardware (IvyBridge i5-3320M)
to figure out the tlb miss costs:

ocperf.py stat -e dtlb_load_misses.walk_duration,dtlb_load_misses.walk_completed,dtlb_store_misses.walk_duration,dtlb_store_misses.walk_completed,itlb_misses.walk_duration,itlb_misses.walk_completed,itlb.itlb_flush

7,720,030,970 dtlb_load_misses_walk_duration [57.13%]
169,856,353 dtlb_load_misses_walk_completed [57.15%]
708,832,859 dtlb_store_misses_walk_duration [57.17%]
19,346,823 dtlb_store_misses_walk_completed [57.17%]
2,779,687,402 itlb_misses_walk_duration [57.15%]
82,241,148 itlb_misses_walk_completed [57.13%]
770,717 itlb_itlb_flush [57.11%]

Show that a dtlb miss is 17.1ns (~45 cycles) and a itlb miss is 13.0ns
(~34 cycles). At those rates, refilling the 512-entry dTLB takes
22,000 cycles. On a SandyBridge system with more cores and larger
caches, those are dtlb=13.4ns and itlb=9.5ns.

Intuition here is that invalidate caused refills will almost always be serviced from the L2
or better since we've recently walked to modify the page needing flush and thus pre-warmed the caches
for any refill? Or is this an artifact of the flush/refill test setup? Main mem latency even on Ivybridge is ~100
clocks, worse in previous generations, so to get down to average ~30 cycle refill you basically can never be
missing in the L1 or maybe L2 which seems optimistic.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/