adding a fetch test to perf tools #193
Conversation
|
Needs rebase now |
|
done |
| return EXIT_FAILURE; | ||
| } | ||
|
|
||
| ctx = OSSL_LIB_CTX_new(); |
There was a problem hiding this comment.
I wonder if we should have a "warm up" option that would do one fetch here outside the measurements. The initial fetch on a libctx with no providers loaded will load the default provider implicitly and fighting for that could skew the numbers fairly visibly. I.e. it would be interesting to see the differences in the numbers. Or instead we could explicitly load the default provider with OSSL_PROVIDER_load().
There was a problem hiding this comment.
I suppose that would depend on the case we want to measure. I could envision wanting to measure both the fast path (where a fetch would always hit in the cache), and the slow path (where a fetch would never hit in the cache)
Can we save that for a subsequent PR?
There was a problem hiding this comment.
Another PR I think.
IMO a warm up isn't going to impact the measurement significantly given a large number of iterations per thread.
| } | ||
|
|
||
| if (fetch_type != NULL) { | ||
| exclusive_fetch_alg = strstr(fetch_type, ":"); |
There was a problem hiding this comment.
strchr()? I think strstr() is a big hammer here.
perf/evp_fetch.c
Outdated
| #include <openssl/core_names.h> | ||
| #include "perflib/perflib.h" | ||
|
|
||
| #define NUM_CALLS_PER_TEST 100000 |
There was a problem hiding this comment.
Please consider raising this number. Even 10000000 calls takes just a few seconds on my laptop.
|
Merged. Thank you for your contribution. |
Reviewed-by: Paul Dale <ppzgs1@gmail.com> Reviewed-by: Tomas Mraz <tomas@openssl.org> (Merged from openssl#193)
Reviewed-by: Paul Dale <ppzgs1@gmail.com> Reviewed-by: Tomas Mraz <tomas@openssl.org> (Merged from openssl#193)
Reviewed-by: Paul Dale <ppzgs1@gmail.com> Reviewed-by: Tomas Mraz <tomas@openssl.org> (Merged from openssl#193)
Add a perf test to record how many EVP fetches we can do in unit time