Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Performance vs. Prometheus SDK #5542

Open
yurishkuro opened this issue Jun 25, 2024 · 6 comments
Open

Performance vs. Prometheus SDK #5542

yurishkuro opened this issue Jun 25, 2024 · 6 comments
Labels
area:metrics Part of OpenTelemetry Metrics enhancement New feature or request

Comments

@yurishkuro
Copy link
Member

Description

Jaeger is in the process of migrating away from Prometheus SDK towards OTEL SDK. We're currently blocked by a massive performance degradation, as illustrated by this benchmark jaegertracing/jaeger#5676. Are we not using OTEL SDK correctly? We're seeing 10-25x slowdown compared to Prometheus SDK.

$ go test -benchmem -benchtime=2s -bench=Benchmark ./internal/metrics/
BenchmarkPrometheusCounter-10       	342003924	         6.984 ns/op	       0 B/op	       0 allocs/op
BenchmarkOTELCounter-10             	33299455	        71.73 ns/op	       0 B/op	       0 allocs/op
BenchmarkOTELCounterWithLabel-10    	12442818	       190.6 ns/op	      16 B/op	       1 allocs/op

Environment

  • OS: macOS
  • Architecture: arm64
  • Go Version: 1.22
  • opentelemetry-go version: 1.27.0

Steps To Reproduce

jaegertracing/jaeger#5676

Expected behavior

Expecting to see counter bumps to be in the ballpark with Prometheus counters.
@yurishkuro yurishkuro added the area:metrics Part of OpenTelemetry Metrics label Jun 25, 2024
@pellared
Copy link
Member

For me it looks like a correct usage of OTel Metrics SDK.

@pellared pellared added enhancement New feature or request pkg:exporter:prometheus Related to the Prometheus exporter package labels Jun 25, 2024
@github-project-automation github-project-automation bot moved this to Needs triage in Go: Triage Jun 25, 2024
@pellared pellared removed this from Go: Triage Jun 25, 2024
@dashpole dashpole mentioned this issue Jun 25, 2024
Closed
@dashpole
Copy link
Contributor

dashpole commented Jun 25, 2024 •
edited
Loading

To summarize my findings in #5544.

  • We should take a close look at the exemplar reservoir performance when exemplars are disabled. It currently makes up a substantial (~50%) portion of the overhead for the no-attributes case.
  • OTel could potentially consider bound instruments if we want to achieve performance similar to the prometheus client. Bound instruments would potentially eliminate ~95% of the current overhead (excluding overhead from exemplar recording) for the with-attributes case.
  • For a very small improvement, we could consider optimizing the code which increments our counters values similar to what prometheus has done. This would be more noticeable if we implemented other the optimizations above.

@yurishkuro
Copy link
Member Author

Bound instruments OTEP https://github.com/open-telemetry/oteps/blob/main/text/metrics/0070-metric-bound-instrument.md

@dashpole
Copy link
Contributor

We should take a close look at the exemplar reservoir performance when exemplars are disabled. It currently makes up a substantial (~50%) portion of the overhead for the no-attributes case.

This appears to be because of the time.Now() call for each measurement. We should at least consider moving the time.Now call into the exemplar reservoir so that it is only invoked when we are actually recording an exemplar.

@dashpole
Copy link
Contributor

I also found that the benchmark did not change if I swapped out the OTel prometheus exporter with a manual reader (which is expected). I'm removing the prometheus exporter label.

@dashpole dashpole removed the pkg:exporter:prometheus Related to the Prometheus exporter package label Jun 25, 2024
@dashpole dashpole mentioned this issue Jun 25, 2024
Merged
@dashpole
Copy link
Contributor

#5545 is a ~45% performance improvement for the zero-attributes case, and a ~20% performance improvement for the single-attribute case.

@XSAM XSAM mentioned this issue Jun 26, 2024
Open
4 tasks
MrAlias added a commit that referenced this issue Jul 1, 2024
@dashpole @XSAM @MrAlias
Move time.Now call into exemplar reservoir to improve performance (#5545
0485de2 
)

Part of addressing
#5542.

### Motivation

This removes the `time.Now()` call from filtered-out Exemplars by only
invoking `time.Now()` after the filtering decision is made. This
improvement is especially noticeable for measurements without any
attributes.

```
goos: linux
goarch: amd64
pkg: go.opentelemetry.io/otel/sdk/metric
cpu: AMD EPYC 7B12
                                                                 │   old.txt    │                new.txt                │
                                                                 │    sec/op    │   sec/op     vs base                  │
SyncMeasure/NoView/Int64Counter/Attributes/0-24                    158.20n ± 4%   99.83n ± 1%  -36.90% (p=0.000 n=10)
SyncMeasure/NoView/Int64Counter/Attributes/1-24                     333.3n ± 4%   274.8n ± 1%  -17.55% (p=0.000 n=10)
SyncMeasure/NoView/Int64Counter/Attributes/10-24                    1.640µ ± 1%   1.600µ ± 1%   -2.41% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/0-24                   159.0n ± 3%   101.3n ± 0%  -36.27% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/1-24                   340.0n ± 2%   272.0n ± 1%  -20.00% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/10-24                  1.661µ ± 1%   1.597µ ± 0%   -3.85% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/0-24               159.8n ± 1%   103.1n ± 0%  -35.50% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/1-24               339.5n ± 1%   273.1n ± 0%  -19.57% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/10-24              1.656µ ± 0%   1.589µ ± 0%   -4.05% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/0-24             159.3n ± 2%   100.8n ± 0%  -36.74% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/1-24             337.9n ± 2%   271.8n ± 1%  -19.55% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/10-24            1.657µ ± 0%   1.593µ ± 1%   -3.83% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/0-24                  144.65n ± 4%   89.38n ± 0%  -38.21% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/1-24                   235.7n ± 2%   183.5n ± 0%  -22.15% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/10-24                  900.8n ± 1%   836.8n ± 0%   -7.10% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/0-24                145.60n ± 5%   93.48n ± 1%  -35.80% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/1-24                 240.9n ± 1%   183.0n ± 0%  -24.06% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/10-24                905.6n ± 1%   826.3n ± 0%   -8.76% (p=0.000 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/0-24                   20.33n ± 0%   20.35n ± 0%        ~ (p=0.302 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/1-24                   26.46n ± 0%   26.45n ± 1%        ~ (p=0.868 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/10-24                  26.50n ± 0%   26.47n ± 0%        ~ (p=0.208 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/0-24                 20.34n ± 1%   20.27n ± 0%   -0.34% (p=0.009 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/1-24                 26.55n ± 0%   26.60n ± 1%        ~ (p=0.109 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/10-24                26.59n ± 1%   26.57n ± 1%        ~ (p=0.926 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/0-24             20.38n ± 1%   20.38n ± 0%        ~ (p=0.725 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/1-24             26.39n ± 0%   26.44n ± 0%        ~ (p=0.238 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/10-24            26.52n ± 0%   26.42n ± 0%   -0.36% (p=0.049 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/0-24           20.30n ± 0%   20.25n ± 0%        ~ (p=0.196 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/1-24           26.57n ± 0%   26.54n ± 1%        ~ (p=0.540 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/10-24          26.57n ± 0%   26.51n ± 1%        ~ (p=0.643 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/0-24                 20.37n ± 0%   20.36n ± 1%        ~ (p=1.000 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/1-24                 26.41n ± 0%   26.50n ± 0%   +0.32% (p=0.007 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/10-24                26.44n ± 0%   26.55n ± 1%   +0.42% (p=0.012 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/0-24               20.30n ± 0%   20.45n ± 0%   +0.74% (p=0.000 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/1-24               26.52n ± 0%   26.48n ± 0%        ~ (p=0.127 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/10-24              26.55n ± 0%   26.48n ± 0%   -0.26% (p=0.002 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/0-24             170.5n ± 2%   110.8n ± 0%  -35.03% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/1-24             402.5n ± 1%   331.5n ± 1%  -17.64% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/10-24            1.363µ ± 1%   1.281µ ± 1%   -6.02% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/0-24           170.6n ± 1%   111.5n ± 1%  -34.64% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/1-24           397.1n ± 1%   335.9n ± 0%  -15.41% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/10-24          1.371µ ± 1%   1.279µ ± 1%   -6.71% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/0-24       170.1n ± 1%   112.2n ± 0%  -34.09% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/1-24       397.5n ± 1%   330.2n ± 0%  -16.93% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/10-24      1.371µ ± 1%   1.289µ ± 1%   -5.95% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/0-24     171.4n ± 2%   112.9n ± 0%  -34.13% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/1-24     397.0n ± 3%   336.4n ± 0%  -15.24% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/10-24    1.383µ ± 1%   1.305µ ± 1%   -5.61% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Histogram/Attributes/0-24          157.30n ± 2%   98.58n ± 1%  -37.33% (p=0.000 n=6+10)
```

### Changes

* Introduce `exemplar.Filter`, which is a filter function based on the
context. It will not be user-facing, so we can always add other
parameters later if needed.
* Introduce `exemplar.FilteredReservoir`, which is similar to a
reservoir, except it does not receive a timestamp. It gets the current
time after the filter decision has been made. It uses generics to avoid
the call to exemplar.NewValue(), since it is internal-only.
* The `exemplar.Reservoir` is left as-is, so that it can be made public
when exemplars are stable. It still includes a timestamp argument.
* Unit tests are updated to expect a much lower number of calls to
time.Now
* `exemplar.Drop` is now an `exemplar.FilteredReservoir` instead of a
`Reservoir`, since we don't need a Reservoir to store things in if the
measurement is always dropped.

Co-authored-by: Sam Xie <[email protected]>
Co-authored-by: Tyler Yahn <[email protected]>
@MrAlias MrAlias mentioned this issue Jul 11, 2024
Closed
6 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
area:metrics Part of OpenTelemetry Metrics enhancement New feature or request
Projects
Go: Metric Prometheus Exporter (GA)
Status: No status
Milestone
No milestone
Development

No branches or pull requests
3 participants
@dashpole @yurishkuro @pellared