using `runtime.Gosched` requires acquiring global lock to check if there are any other goroutines to perform tasks. with the latest versions of runtime it can pause running goroutines automatically without requiring to call `Gosched` directly.
Updates #3966
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
Call runtime.Gosched() only when there is a work to steal from other workers.
Simplify the timeseriesWorker() and unpackWroker() code a bit by inlining stealTimeseriesWork() and stealUnpackWork().
This should reduce CPU usage when processing queries on systems with big number of CPU cores.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3966
Previously the selected time series were split evenly among available CPU cores
for further processing - e.g unpacking the data and applying the given rollup
function to the unpacked data.
Some time series could be processed slower than others.
This could result in uneven work distribution among available CPU cores,
e.g. some CPU cores could complete their work sooner than others.
This could slow down query execution.
The new algorithm allows stealing time series to process from other CPU cores
when all the local work is done. This should reduce the maximum time
needed for query execution (aka tail latency).
The new algorithm should also scale better on systems with many CPU cores,
since every CPU processes locally assigned time series without inter-CPU communications.
The inter-CPU communications are used only when all the local work is finished
and the pending work from other CPUs needs to be stealed.
Unpack time series with less than 400K samples in the currently running goroutine.
Previously a new goroutine was being started for unpacking the samples.
This was requiring additional memory allocations.
Usually the number of blocks returned per each time series during queries is around 4.
So it is a good idea to pre-allocate 4 block references per time series
in order to reduce the number of memory allocations.
Reduce inter-CPU communications when processing the query over big number of time series.
This should improve performance for queries over big number of time series
on systems with many CPU cores.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2896
Based on b596ac3745
Thanks to @zqyzyq for the idea.
- Use binary search instead of linear scan when locating the run of smallest timestamps
in blocks with intersected time ranges. This should improve performance
when merging blocks with big number of samples
- Skip samples with duplicate timestamps. This should increase query performance
in cluster version of VictoriaMetrics with the enabled replication.
Previously SearchMetricNames was returning unmarshaled metric names.
This wasn't great for vmstorage, which should spend additional CPU time
for marshaling the metric names before sending them to vmselect.
While at it, remove possible duplicate metric names, which could occur when
multiple samples for new time series are ingested via concurrent requests.
Also sort the metric names before returning them to the client.
This simplifies debugging of the returned metric names across repeated requests to /api/v1/series
querytracer has been added to the following storage.Storage methods:
- RegisterMetricNames
- DeleteMetrics
- SearchTagValueSuffixes
- SearchGraphitePaths
This allows filling the seriesCountByFocusLabelValue list in the /api/v1/status/tsdb response
with label values for the specified focusLabel, which contain the highest number of time series.
TODO: add this to Cardinality explorer at VMUI - https://docs.victoriametrics.com/#cardinality-explorer
* Rename -search.maxMetricsPointSearch to -search.maxSamplesPerQuery, so it is more consistent with the existing -search.maxSamplesPerSeries
* Move the -search.maxSamplesPerQuery from vmstorage to vmselect, so it could effectively limit the number of raw samples obtained from all the vmstorage nodes
* Document the -search.maxSamplesPerQuery in docs/CHANGELOG.md
