Commit graph

116 commits

Author SHA1 Message Date
Dmytro Kozlov
5256af2291
lib/mergeset: fix linter error () 2022-07-13 12:34:28 +03:00
Aliaksandr Valialkin
7cbcbea49d
lib/mergeset: optimize merge speed a bit
Use heap.Fix instead of heap.Pop + heap.Push when merging blocks
2022-07-12 12:52:36 +03:00
Loki's Wager
ca4730c00f
BugFix part_header.go ()
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2757

Co-authored-by: haotingyi <haotingyi@corp.netease.com>
2022-06-21 15:59:11 +03:00
Aliaksandr Valialkin
fedfc9e686
lib/storage: stop background merge when storage enters read-only mode
This should prevent from `no space left on device` errors when VictoriaMetrics
under-estimates the additional disk space needed for background merge.

See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2603
2022-06-01 14:22:12 +03:00
Aliaksandr Valialkin
ed1b394a1a
app/vmstorage: expose vm_indexdb_items_added_total and vm_indexdb_items_added_size_bytes_total counters at /metrics page
These counters can be used for monitoring the rate of addition of new entries in indexdb (aka inverted index).

See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2471
2022-04-21 13:19:42 +03:00
Aliaksandr Valialkin
edb139cfe4
lib/memory: export process_memory_limit_bytes metric, which shows the amounts of memory the current process has access to
This metric is equivalent to `vm_available_memory_bytes`, but it has better name,
since the metric is related to a process, not VictoriaMetrics itself.

Leave `vm_available_memory_bytes` for backwards compatibility.
2022-04-07 15:24:08 +03:00
Aliaksandr Valialkin
8ef9348801
lib/mergeset: use more rawItemsShard shards on multi-CPU systems
This should improve the scalability for registering of new time series on multi-CPU system
2022-04-06 19:50:41 +03:00
Aliaksandr Valialkin
db00ddd23e
lib/mergeset: skip common prefixes when comparing inmemoryBlock items
This should improve the performance for items sorting inside inmemoryBlock.MarshalUnsortedData
if they have common prefix.

While at it, improve the performance for inmemoryBlock.updateCommonPrefix for sorted items.
This should improve performance for inmemoryBlock.MarshalSortedData during background merge.
2022-04-06 18:55:25 +03:00
Aliaksandr Valialkin
3f999b11f7
lib/mergeset: remove aux buffers from inmemoryPart
This should reduce the size of inmemoryPart items and may improve performance a bit during registering new time series

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2247
2022-03-03 17:12:25 +02:00
Aliaksandr Valialkin
ecf68da79e
lib/mergeset: eliminate copying of itemsData and lensData from storageBlock to inmemoryBlock
This should improve performance when registering new time series.

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2247
2022-03-03 17:12:25 +02:00
Aliaksandr Valialkin
ecf4f7bf21
lib/mergeset: consistency renaming: ip->mp for inmemoryPart vars 2022-03-03 17:12:25 +02:00
Aliaksandr Valialkin
f4e466955d
lib/mergeset: move storageBlock from inmemoryPart to a sync.Pool
The lifetime of storageBlock is much shorter comparing to the lifetime of inmemoryPart,
so sync.Pool usage should reduce overall memory usage and improve performance
because of better locality of reference when marshaling inmemoryBlock to inmemoryPart.

https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2247
2022-03-03 17:12:25 +02:00
Aliaksandr Valialkin
b47f18f555
lib/{mergeset,storage}: tune compression levels for small blocks
This should reduce CPU usage spent on compression
2022-02-25 15:34:13 +02:00
Aliaksandr Valialkin
1967b9c211
lib/mergeset: remove superflouos sorting of inmemoryBlock.data at inmemoryBlock.sort()
There is no need to sort the underlying data according to sorted items there.
This should reduce cpu usage when registering new time series in `indexdb`.

Thanks to @ahfuzhang for the suggestion at https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2245
2022-02-24 11:19:44 +02:00
Aliaksandr Valialkin
986c12487a
lib/mergeset: typo fix after b6ed9afd6d 2022-02-21 19:00:04 +02:00
Roman Khavronenko
bd7837d524
lib: allow to configure cache size by type ()
* lib: allow to configure cache size by type

https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1940
Signed-off-by: hagen1778 <roman@victoriametrics.com>

* Apply suggestions from code review

* wip

Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-21 13:55:51 +02:00
Aliaksandr Valialkin
895c9f4f11
lib/mergeset: tune indexdb/{indexBlocks,dataBlocks} cache sizes further according to production stats
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-02-10 19:11:13 +02:00
Aliaksandr Valialkin
a0a56d6c1c
lib/mergeset: tune sizes for indexdb/dataBlocks and indexdb/indexBlocks according to production workload
This should help with https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007#issuecomment-1032308742
2022-02-08 18:04:03 +02:00
Aliaksandr Valialkin
d8d59ff760
lib/mergeset: pre-allocate data and items for inmemoryBlock in order to reduce memory allocations under high churn rate
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-02-01 11:20:20 +02:00
Aliaksandr Valialkin
02b2bfcff3
lib/bytesutil: split Resize* funcs to MayOverallocate and NoOverallocate for more fine-grained control over memory allocations
Follow-up for f4989edd96

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-02-01 11:20:20 +02:00
匠心零度
a1083d0531
optimized code ()
* optimized code ,because only the first error,so no need var errors []error

* optimized code ,because only the first error,so no need var errors []error

Co-authored-by: lirenzuo <lirenzuo@shein.com>
2022-01-28 12:10:47 +02:00
Aliaksandr Valialkin
6232eaa938
lib/bytesutil: split Resize() into ResizeNoCopy() and ResizeWithCopy() functions
Previously bytesutil.Resize() was copying the original byte slice contents to a newly allocated slice.
This wasted CPU cycles and memory bandwidth in some places, where the original slice contents wasn't needed
after slize resizing. Switch such places to bytesutil.ResizeNoCopy().

Rename the original bytesutil.Resize() function to bytesutil.ResizeWithCopy() for the sake of improved readability.

Additionally, allocate new slice with `make()` instead of `append()`. This guarantees that the capacity of the allocated slice
exactly matches the requested size. The `append()` could return a slice with bigger capacity as an optimization for further `append()` calls.
This could result in excess memory usage when the returned byte slice was cached (for instance, in lib/blockcache).

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-01-25 15:28:42 +02:00
Aliaksandr Valialkin
7ec0705b98
lib/mergeset: allocate the needed amounts of memory when unmarshaling inmemoryBlock
This should reduce the memory required for indexdb/dataBlocks cache.

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-01-24 18:52:22 +02:00
Aliaksandr Valialkin
9edf407144
lib/mergeset: tune caches size limits for indexdb/dataBlocks and indexdb/indexBlocks
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-01-21 12:46:05 +02:00
Aliaksandr Valialkin
6ae584b9b3
lib/{mergeset,storage}: properly limit cache sizes for indexdb
Previously these caches could exceed limits set via `-memory.allowedPercent` and/or `-memory.allowedBytes`,
since limits were set independently per each data part. If the number of data parts was big, then limits could be exceeded,
which could result to out of memory errors.

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
2022-01-20 18:45:03 +02:00
Aliaksandr Valialkin
ab4be24397
app/vmstorage: export vm_cache_size_max_bytes metrics for determining capacity of various caches
The vm_cache_size_max_bytes metric can be used for determining caches which reach their capacity via the following query:

   vm_cache_size_bytes / vm_cache_size_max_bytes > 0.9
2021-12-02 10:30:01 +02:00
Aliaksandr Valialkin
b885bd9b7d lib/{mergeset,storage}: improve the detection of the needed free space for background merge
This should prevent from possible out of disk space crashes during big merges.

See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1560
2021-08-25 10:01:09 +03:00
Aliaksandr Valialkin
51cd19d2e3 lib/storage: reset perKeyMisses stats less frequently
This should reduce CPU usage for queries executed with intervals higher than 30 seconds
2021-07-12 14:34:54 +03:00
Aliaksandr Valialkin
41754e12f8 lib/mergeset: cache indexBlock items only on the second request
This should reduce the indexdb/indexBlocks cache size, since it won't contain one-time-wonders items.
2021-07-07 15:24:37 +03:00
Aliaksandr Valialkin
9826f7c1be lib/storage: do not cache inmemoryBlock entries requested only once (aka one-time-wonder items)
This should reduce the cache size and memory usage for the indexdb/dataBlocks cache
2021-07-07 10:59:45 +03:00
Aliaksandr Valialkin
b805a675f3 lib/{mergeset,storage}: switch from sync.Pool to chan-based pool for inmemoryPart objects
This should reduce memory usage on systems with big number of CPU cores,
since every inmemoryPart object occupies at least 64KB of memory and sync.Pool maintains
a separate pool inmemoryPart objects per each CPU core.

Though the new scheme for the pool worsens per-cpu cache locality, this should be amortized
by big sizes of inmemoryPart objects.
2021-07-06 16:33:25 +03:00
Aliaksandr Valialkin
db6bd69475 lib/mergeset: increase pool capacity for inmemoryBlock according to collected profiles from production workload
CPU and memory profiles show that the pool capacity for inmemoryBlock objects is too small.
This results in the increased load on memory allocation code in Go runtime.
Increase the pool capacity in order to reduce the load on Go runtime.
2021-07-06 13:44:27 +03:00
Aliaksandr Valialkin
fd32855a6c lib/mergeset: limit the frequency for flushCallback calls to once per 10 seconds
This should improve hit ratio for tagFiltersCache when big number of new time series are constantly registered
(aka high churn rate). This, in turn, should reduce CPU usage for queries over such time series.
2021-07-06 12:20:15 +03:00
Aliaksandr Valialkin
f12f97daa1 lib/{storage,mergeset}: increase cache timeout for data and index blocks from a minute to two minutes
One minute cache timeout result in slower queries in some production workloads where the interval
between query execution is in the range 1 minute - 2 minutes.
2021-07-05 14:25:59 +03:00
Aliaksandr Valialkin
8055439fe4 lib/storage: properly detect free disk space shortage during data merge
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1373
2021-07-02 17:42:23 +03:00
Aliaksandr Valialkin
bced9ee666 lib/{mergeset,storage}: reduce the maximum lifetime for cached indexdb and data blocks from 2 minutes to a minute
This should reduce memory usage on a system with high number of active time series and a high churn rate.
One minute is enough for caching the blocks needed for repeated queries (e.g. alerting rules, recording rules and dashboard refreshes).
2021-06-29 19:57:53 +03:00
Aliaksandr Valialkin
b7c0b3dde3 lib/mergeset: switch from sync.Pool to a channel for a pool for inmemoryBlock structs
This should reduce memory usage for the pool on systems with big number of CPU cores.

The sync.Pool maintains per-CPU pools, so the total number of objects in the pool
is proportional to the number of available CPU cores. The channel limits the number
of pooled objects by its own capacity. This means smaller number of pooled objects on average.
2021-06-29 19:57:52 +03:00
Aliaksandr Valialkin
0efd37cec1 lib/{mergeset,storage}: reduce the number of fsync calls on data ingestion path on systems with many cpu cores
VictoriaMetrics maintains a buffer per CPU core for the ingested data. These buffers are flushed to disk every second.
These buffers are flushed to disk in parallel starting from the commit 56b6b893ce .
This resulted in increased write disk IO usage on systems with many cpu cores
as described at https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1338#issuecomment-863046999 .

This commit merges the per-CPU buffers into bigger in-memory buffers before flushing them to disk.
This should reduce the rate of fsync syscalls and, consequently, the write disk IO on systems with many CPU cores.

This should help https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1338
See also https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1244
2021-06-17 13:51:42 +03:00
Aliaksandr Valialkin
6865f3b497 Revert "lib/mergeset: remove a pool for inmemoryBlock structs"
This reverts commit 793fe39921.

Reason to revert: production testing revealed possible slowdown when registering big number of new time series
2021-05-28 01:11:22 +03:00
Aliaksandr Valialkin
7b33bc67a1 lib/mergeset: remove a pool for inmemoryBlock structs
The pool for inmemoryBlock struct doesn't give any performance gains in production workloads,
while it may result in excess memory usage for inmemoryBlock structs inside the pool during
background merge of indexdb.
2021-05-27 22:00:50 +03:00
Aliaksandr Valialkin
0fc857d363 lib/{mergeset,storage}: reduce the number of IFNO log messages like merged ... items across ... blocks in ... seconds
Log these messages if the merge takes more than 30 seconds instead of 10 seconds.
2021-05-23 14:15:49 +03:00
Aliaksandr Valialkin
e37e1b1e34 lib/{storage,mergeset}: fix unaligned 64-bit atomic operation panic for 32-bit architectures
The panic has been introduced in 56b6b893ce
2021-04-27 16:42:19 +03:00
Aliaksandr Valialkin
2d1d60118d lib/mergeset: split rows ingestion among multiple shards
This improves rows ingestion on systems with many CPU cores by reducing lock contention.

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1244

Thanks to @waldoweng for the original idea and draft implementation at https://github.com/VictoriaMetrics/VictoriaMetrics/pull/1243
2021-04-27 15:45:11 +03:00
Aliaksandr Valialkin
ab8008d6d7 lib/{storage,mergeset}: remove empty directories on startup. Such directories can be left after unclean shutdown on NFS storage
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1142
2021-04-22 13:03:29 +03:00
Aliaksandr Valialkin
587132555f lib/mergeset: reduce memory usage for inmemoryBlock by using more compact items representation
This also should reduce CPU time spent by GC, since inmemoryBlock.items don't have pointers now,
so GC doesn't need visiting them.
2021-02-21 22:09:10 +02:00
Aliaksandr Valialkin
b8a5ee2e93 lib/{mergeset,storage}: allow merging smaller number of small parts
While this may increase CPU and disk IO usage needed for background merge,
this also recudes CPU usage during queries in production. This is because
such queries tend to read recently added data and it is better to have lower number
of parts for such data in order to reduce CPU usage.

This partially reverts ebf8da3730
2021-02-21 21:43:37 +02:00
Aliaksandr Valialkin
34195218e1 lib/{mergeset,storage}: do not use pools for indexBlock and inmemoryBlock during their caching, since this results in higher memory usage in production without any performance gains 2021-02-21 21:43:37 +02:00
Aliaksandr Valialkin
9566015a36 Revert "lib/mergeset: tune lifetime for entries inside block caches"
This reverts commit 458c89324d.

Production testing revealed zero improvements for memory usage with reduced lifetime for entries in block caches.
2021-02-17 20:42:15 +02:00
Aliaksandr Valialkin
35a23234ca lib/mergeset: tune lifetime for entries inside block caches
This should reduce memory usage in general case without significant CPU usage increase
2021-02-16 18:12:32 +02:00
Aliaksandr Valialkin
500acb958c lib/mergeset: clarify comments in the code a bit 2021-02-16 18:03:11 +02:00