[CHANGED] sub benchmark tuned #777
Conversation
levb
commented
Aug 5, 2024
levb
commented
- Added CI steps to compare to the base branch
- Dialed down # messages, tuned the matrix
- Added locking in the bench to clear sanitize=thread
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #777 +/- ##
==========================================
+ Coverage 68.71% 68.83% +0.12%
==========================================
Files 39 39
Lines 15207 15251 +44
Branches 3143 3157 +14
==========================================
+ Hits 10449 10498 +49
+ Misses 1700 1682 -18
- Partials 3058 3071 +13 ☔ View full report in Codecov by Sentry. |
| @@ -222,6 +230,7 @@ static void _benchMatrix(threadConfig *threadsVector, int lent, int *subsVector, | |||
| } | |||
There was a problem hiding this comment.
Looks like you write the number of messages as numPubMessages * numSubs, where numPubMessages is NMessages / numSubs. So you should not simply write NMessages? Because of rounding of the first division, the number of messages may not be what you expect after the multiplication.
There was a problem hiding this comment.
This is so that each sub gets the same number of messages, since we average/best/worst across subs. The output includes the actual total number of messages received.
There was a problem hiding this comment.
But again, we have:
int numPubMessages = NMessages / numSubs;
...
printf("\t{\"subs\":%d, \"threads\":%d, \"messages\":%d, \"best\":%d, \"average\":%d, \"worst\":%d}%s\n",
numSubs, env->threads.useGlobalDelivery ? env->threads.max : 0, numPubMessages * numSubs, best, average, worst, comma);
This is for a given _benchMatrix run. So we divide the number of messages (NMessages) by the number of subs, but then print out that number * numSubs. So to me, it looks like using NMessages should be used?
There was a problem hiding this comment.
Oh right, sorry I missed the question. "messages" is included just to be part of the key, it is not used in any calculations. For the purpose of the key (combined with NSubs) it's as good as NMessages, but it was useful for me to see it to visually confirm the correctness of how it was calculated.
|
@kozlovic I answered the questions, not sure there are any changes to make? |
|
@kozlovic I apologize if I confused by mentioning it in the "child" PR, it's this one I need your LGTM on, otherwise I'd need to bypass the branch protection, which I'd rather not. |
|
@levb But on that one, I still had a comment that I don't think you addressed: https://github.com/nats-io/nats.c/pull/777/files#r1706002494? |
