Speed up JSON output #1114
Speed up JSON output #1114
Conversation
The commit includes three main changes: 1. Don't write on Collect instead (as all other outputs) append to a buffer and write the buffer in a separate goroutine every 100 milliseconds 2. Use json.Encoder for speed up 3. If the output file ends in `.gz` gzip the output which in my testing decreases the file size by 30x. I couldn't measure any performance degradation from the fact we need to compress if anything there is some performance gain because of the smaller io writes I was testing with simple script doing batch requests for a local zero length file with 40VUS, that makes around 8k RPS using around 600-700mb of memory. When running before those I was getting 5.8-6k RPS, after this I am getting 6.5-6.9k RPS with around 3GB of memory usage. At this point my cpu and memory profiling shows that the problem is the json encoding upwards of 10 seconds and as such the next possible optimizations are in using some kind of faster json encoding.
|
I would like to add that the test here and in #1113 is basically using all available CPU either way ... so there is essentially no head room so adding an output WILL decrease performance in all cases. In both PRs I am basically making that hit smaller and hopefully less noticeable in the cases where the CPU usage is at 100% , while still saving all of the metric data. In this PR for example the fact that the gzip decreases the json output by 30x makes the json output viable as a metric output even for large and long tests. The example above with 6.5k+ RPS for 4 minutes was producing 3GB files which got reduced to 100-120MB with gzip. I did some testing and around 5.5-5.6RPS all the json writes finish within 100ms or less which makes it viable for that load IMO |
Codecov Report
@@ Coverage Diff @@
## master #1114 +/- ##
==========================================
- Coverage 73.25% 73.09% -0.16%
==========================================
Files 141 141
Lines 10260 10282 +22
==========================================
Hits 7516 7516
- Misses 2302 2322 +20
- Partials 442 444 +2
Continue to review full report at Codecov.
|
The commit includes three main changes:
buffer and write the buffer in a separate goroutine every 100
milliseconds
.gzgzip the output which in my testingdecreases the file size by 30x. I couldn't measure any performance
degradation from the fact we need to compress if anything there is some
performance gain because of the smaller io writes
I was testing with simple script doing batch requests for a local zero
length file with 40VUS, that makes around 8k RPS using around 600-700mb
of memory. When running before those I was getting 5.8-6k RPS, after
this I am getting 6.5-6.9k RPS with around 3GB of memory usage.
At this point my cpu and memory profiling shows that the problem is the
json encoding upwards of 10 seconds and as such the next possible
optimizations are in using some kind of faster json encoding.