Hide OpenSSL sha256 initialisation in a function-scoped `thread_local` #7254

eddyashton · 2025-09-04T15:15:59Z

See discussion in this thread:
https://github.com/microsoft/CCF/pull/7251/files#r2322012456

I think this is much nicer, but it does seem to be slightly slower (checking a bool for thread_local initialisation on every call), so this is debatable.

eddyashton · 2025-09-04T15:23:38Z

src/crypto/test/bench.cpp

+PICOBENCH_SUITE("digest sha256");
+namespace SHA256_bench
+{
+  const std::vector<int> sha256_shifts = {
+    2 << 4, 2 << 6, 2 << 8, 2 << 10, 2 << 12, 2 << 14, 2 << 16};
+
+  auto openssl_sha256_preinit = sha256_bench;
+  PICOBENCH(openssl_sha256_preinit).iterations(sha256_shifts).baseline();

-DEFINE_SHA256_BENCH(6)
-DEFINE_SHA256_BENCH(8)
-DEFINE_SHA256_BENCH(10)
-DEFINE_SHA256_BENCH(12)
-DEFINE_SHA256_BENCH(14)
-DEFINE_SHA256_BENCH(16)
+  auto openssl_sha256_tl_init = sha256_bench_;
+  PICOBENCH(openssl_sha256_tl_init).iterations(sha256_shifts);
+
+  auto openssl_sha256_nocache = sha256_noopt_bench;
+  PICOBENCH(openssl_sha256_nocache).iterations(sha256_shifts);
+}


This refactor is worth doing in a separate, even if we don't use this initialisation. It moves all these benchmarks into a single suite, and uses the .iterations (the Dim column in the output) for the actual digest size, so that the ns/op and Ops/second columns are actually talking about per-byte costs, rather than some abstract multiple.

Before:

## digest sha256 (2 << 6): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_6_no * | 10 | 0.004 | 442 | - | 2259887.0 openssl_sha256_6 | 10 | 0.003 | 279 | 0.633 | 3572704.5 ## digest sha256 (2 << 8): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_8_no * | 10 | 0.007 | 708 | - | 1412429.4 openssl_sha256_8 | 10 | 0.005 | 498 | 0.704 | 2005615.7 ## digest sha256 (2 << 10): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_10_no * | 10 | 0.017 | 1657 | - | 603281.9 openssl_sha256_10 | 10 | 0.015 | 1512 | 0.913 | 661025.9 ## digest sha256 (2 << 12): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_12_no * | 10 | 0.056 | 5587 | - | 178980.5 openssl_sha256_12 | 10 | 0.054 | 5418 | 0.970 | 184559.7 ## digest sha256 (2 << 14): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_14_no * | 10 | 0.214 | 21420 | - | 46684.7 openssl_sha256_14 | 10 | 0.212 | 21208 | 0.990 | 47151.1 ## digest sha256 (2 << 16): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_16_no * | 10 | 0.846 | 84614 | - | 11818.3 openssl_sha256_16 | 10 | 0.843 | 84281 | 0.996 | 11865.1 ## digest sha256 (2 << 6): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_6_no * | 10 | 0.004 | 442 | - | 2259887.0 openssl_sha256_6 | 10 | 0.003 | 279 | 0.633 | 3572704.5 ## digest sha256 (2 << 8): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_8_no * | 10 | 0.007 | 708 | - | 1412429.4 openssl_sha256_8 | 10 | 0.005 | 498 | 0.704 | 2005615.7 ## digest sha256 (2 << 10): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_10_no * | 10 | 0.017 | 1657 | - | 603281.9 openssl_sha256_10 | 10 | 0.015 | 1512 | 0.913 | 661025.9 ## digest sha256 (2 << 12): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_12_no * | 10 | 0.056 | 5587 | - | 178980.5 openssl_sha256_12 | 10 | 0.054 | 5418 | 0.970 | 184559.7 ## digest sha256 (2 << 14): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_14_no * | 10 | 0.214 | 21420 | - | 46684.7 openssl_sha256_14 | 10 | 0.212 | 21208 | 0.990 | 47151.1 ## digest sha256 (2 << 16): Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_16_no * | 10 | 0.846 | 84614 | - | 11818.3 openssl_sha256_16 | 10 | 0.843 | 84281 | 0.996 | 11865.1

After:

## digest sha256: Name (* = baseline) | Dim | Total ms | ns/op |Baseline| Ops/second --------------------------|--------:|----------:|--------:|-------:|----------: openssl_sha256_noopt * | 32 | 0.000 | 12 | - | 80200501.3 openssl_sha256_opt | 32 | 0.000 | 5 | 0.439 |182857142.9 openssl_sha256_noopt * | 128 | 0.000 | 3 | - |272921108.7 openssl_sha256_opt | 128 | 0.000 | 2 | 0.557 |490421455.9 openssl_sha256_noopt * | 512 | 0.001 | 1 | - |665799739.9 openssl_sha256_opt | 512 | 0.001 | 1 | 0.700 |951672862.5 openssl_sha256_noopt * | 2048 | 0.002 | 0 | - |1137146030.0 openssl_sha256_opt | 2048 | 0.002 | 0 | 0.890 |1277604491.6 openssl_sha256_noopt * | 8192 | 0.006 | 0 | - |1362608117.1 openssl_sha256_opt | 8192 | 0.006 | 0 | 0.934 |1458170167.3 openssl_sha256_noopt * | 32768 | 0.022 | 0 | - |1481977296.4 openssl_sha256_opt | 32768 | 0.022 | 0 | 1.005 |1475105789.1 openssl_sha256_noopt * | 131072 | 0.088 | 0 | - |1490792870.9 openssl_sha256_opt | 131072 | 0.090 | 0 | 1.028 |1450537289.3

The evidence that these are "the same numbers" is that the ratio in the Baseline column is (roughly, within noise) the same for each size. Just now the other columns are more readable, the Total ms is for one rather than 10, and as mentioned above the "op" unit is "byte".

(I've also added 2 << 4, because I was interested)

(Bonus, you can now run this directly with ./crypto_bench --run-suite="digest sha256", because it's a single suite)

eddyashton · 2025-09-04T15:32:05Z

Is this actually faster?

If we only run a single digest, then the benchmarks say no - this is faster than not caching the contexts, but slower than the explicit _init call (since this approach introduces an implicit atomic read, to check whether the thread_local has been constructed).

## digest sha256:

 Name (* = baseline)      |   Dim   |  Total ms |  ns/op  |Baseline| Ops/second
--------------------------|--------:|----------:|--------:|-------:|----------:
 openssl_sha256_noopt *   |      32 |     0.000 |      11 |      - | 84880636.6
 openssl_sha256_opt       |      32 |     0.000 |       5 |  0.440 |192771084.3
 openssl_sha256_tl        |      32 |     0.000 |       7 |  0.599 |141592920.4
 openssl_sha256_noopt *   |     128 |     0.001 |       5 |      - |184971098.3
 openssl_sha256_opt       |     128 |     0.000 |       2 |  0.382 |484848484.8
 openssl_sha256_tl        |     128 |     0.000 |       2 |  0.403 |458781362.0
 openssl_sha256_noopt *   |     512 |     0.001 |       1 |      - |596041909.2
 openssl_sha256_opt       |     512 |     0.001 |       1 |  0.629 |948148148.1
 openssl_sha256_tl        |     512 |     0.001 |       1 |  0.698 |853333333.3
 openssl_sha256_noopt *   |    2048 |     0.002 |       1 |      - |938158497.5
 openssl_sha256_opt       |    2048 |     0.002 |       0 |  0.770 |1219047619.0
 openssl_sha256_tl        |    2048 |     0.002 |       0 |  0.737 |1272840273.5
 openssl_sha256_noopt *   |    8192 |     0.006 |       0 |      - |1348255431.2
 openssl_sha256_opt       |    8192 |     0.006 |       0 |  0.952 |1416320885.2
 openssl_sha256_tl        |    8192 |     0.006 |       0 |  0.960 |1404423109.9
 openssl_sha256_noopt *   |   32768 |     0.023 |       0 |      - |1434173669.5
 openssl_sha256_opt       |   32768 |     0.023 |       0 |  0.990 |1449013885.2
 openssl_sha256_tl        |   32768 |     0.023 |       0 |  0.991 |1447030249.5
 openssl_sha256_noopt *   |  131072 |     0.090 |       0 |      - |1453239164.9
 openssl_sha256_opt       |  131072 |     0.090 |       0 |  0.999 |1454303371.9
 openssl_sha256_tl        |  131072 |     0.090 |       0 |  1.001 |1451726162.1

But that's slightly artificial, what about the minor "warm cache" (/branch predictor) win we'd expect if we digest multiple things in quick succession (simulated by "calling the function 10 times"). Then this approach is near-identical:

## digest sha256:

 Name (* = baseline)      |   Dim   |  Total ms |  ns/op  |Baseline| Ops/second
--------------------------|--------:|----------:|--------:|-------:|----------:
 openssl_sha256_noopt *   |      32 |     0.003 |     103 |      - |  9685230.0
 openssl_sha256_opt       |      32 |     0.001 |      42 |  0.408 | 23721275.0
 openssl_sha256_tl        |      32 |     0.001 |      44 |  0.433 | 22346368.7
 openssl_sha256_noopt *   |     128 |     0.004 |      34 |      - | 29317453.0
 openssl_sha256_opt       |     128 |     0.002 |      17 |  0.516 | 56863616.2
 openssl_sha256_tl        |     128 |     0.002 |      17 |  0.518 | 56587091.1
 openssl_sha256_noopt *   |     512 |     0.007 |      13 |      - | 75095335.9
 openssl_sha256_opt       |     512 |     0.005 |       9 |  0.721 |104086196.4
 openssl_sha256_tl        |     512 |     0.005 |       9 |  0.727 |103350827.6
 openssl_sha256_noopt *   |    2048 |     0.018 |       8 |      - |116542423.0
 openssl_sha256_opt       |    2048 |     0.015 |       7 |  0.876 |133090720.0
 openssl_sha256_tl        |    2048 |     0.016 |       7 |  0.883 |131958762.9
 openssl_sha256_noopt *   |    8192 |     0.060 |       7 |      - |136688246.7
 openssl_sha256_opt       |    8192 |     0.057 |       7 |  0.959 |142506740.9
 openssl_sha256_tl        |    8192 |     0.057 |       7 |  0.959 |142561300.3
 openssl_sha256_noopt *   |   32768 |     0.228 |       6 |      - |143992758.2
 openssl_sha256_opt       |   32768 |     0.226 |       6 |  0.992 |145198025.5
 openssl_sha256_tl        |   32768 |     0.226 |       6 |  0.994 |144806618.1
 openssl_sha256_noopt *   |  131072 |     0.900 |       6 |      - |145589128.8
 openssl_sha256_opt       |  131072 |     0.900 |       6 |  0.999 |145704360.4
 openssl_sha256_tl        |  131072 |     0.900 |       6 |  1.000 |145615331.2

I think that's close enough that it's worth comparing e2e numbers, so I'll try that.

…ench_cleanup

maxtropets · 2025-09-05T09:52:20Z

Thanks for benchmarking, picking that up in #7251

eddyashton added 2 commits September 4, 2025 14:49

Single suite

75a33f5

Add a benchmark of a thread_local RAII initializer

a1055ce

eddyashton commented Sep 4, 2025

View reviewed changes

eddyashton added 3 commits September 4, 2025 15:39

Nuke openssl_init.h

e8418c0

Tidy build

bfd0b90

Merge branch 'main' of github.com:microsoft/CCF into openssl_sha256_b…

39c3350

…ench_cleanup

eddyashton added the bench-ab label Sep 5, 2025

maxtropets closed this Sep 5, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Hide OpenSSL sha256 initialisation in a function-scoped `thread_local` #7254

Hide OpenSSL sha256 initialisation in a function-scoped `thread_local` #7254

Uh oh!

eddyashton commented Sep 4, 2025

Uh oh!

eddyashton Sep 4, 2025

Uh oh!

eddyashton commented Sep 4, 2025

Uh oh!

maxtropets commented Sep 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Hide OpenSSL sha256 initialisation in a function-scoped thread_local #7254

Hide OpenSSL sha256 initialisation in a function-scoped thread_local #7254

Uh oh!

Conversation

eddyashton commented Sep 4, 2025

Uh oh!

eddyashton Sep 4, 2025

Choose a reason for hiding this comment

Uh oh!

eddyashton commented Sep 4, 2025

Is this actually faster?

Uh oh!

maxtropets commented Sep 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Hide OpenSSL sha256 initialisation in a function-scoped `thread_local` #7254

Hide OpenSSL sha256 initialisation in a function-scoped `thread_local` #7254