Add referenced memory metric #2495

katarzyna-z · 2020-04-16T07:39:51Z

see: https://github.com/brendangregg/wss#wsspl-referenced-page-flag

Introduce map of collectors in containerData
Introduce map of managers in Manager

Signed-off-by: Katarzyna Kujawa katarzyna.kujawa@intel.com

The working set size indicates how much memory a container needs to keep working. It's intrusive metric because collection of metric can influence kernel page reclaim policy and add latency (this is mentioned in documentation). Wss metric is disabled by default.

katarzyna-z · 2020-04-16T11:18:01Z

I'm working on fix for data race indicated in tests:

==================
WARNING: DATA RACE
Read at 0x00c0004d2db8 by goroutine 50:
  encoding/json.sliceEncoder.encode()
      /usr/local/go/src/reflect/value.go:1071 +0x137
  encoding/json.sliceEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:760 +0x7b
  encoding/json.structEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:664 +0x40d
  encoding/json.structEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:635 +0xa0
  encoding/json.ptrEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:810 +0xfc
  encoding/json.ptrEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:805 +0x7b
  encoding/json.arrayEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:791 +0xe3
  encoding/json.arrayEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:784 +0x7b
  encoding/json.sliceEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:765 +0xda
  encoding/json.sliceEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:760 +0x7b
  encoding/json.structEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:664 +0x40d
  encoding/json.structEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:635 +0xa0
  encoding/json.mapEncoder.encode()
      /usr/local/go/src/encoding/json/encode.go:706 +0x36e
  encoding/json.mapEncoder.encode-fm()
      /usr/local/go/src/encoding/json/encode.go:682 +0x7b
  encoding/json.(*encodeState).reflectValue()
      /usr/local/go/src/encoding/json/encode.go:337 +0x93
  encoding/json.(*encodeState).marshal()
      /usr/local/go/src/encoding/json/encode.go:309 +0xcc
  encoding/json.Marshal()
      /usr/local/go/src/encoding/json/encode.go:161 +0x73
  github.com/google/cadvisor/cmd/internal/api.writeResult()
      /go/src/github.com/google/cadvisor/cmd/internal/api/handler.go:126 +0x5d
  github.com/google/cadvisor/cmd/internal/api.(*version1_2).HandleRequest()
      /go/src/github.com/google/cadvisor/cmd/internal/api/versions.go:236 +0x570
  github.com/google/cadvisor/cmd/internal/api.(*version1_3).HandleRequest()
      /go/src/github.com/google/cadvisor/cmd/internal/api/versions.go:272 +0x104
  github.com/google/cadvisor/cmd/internal/api.handleRequest()
      /go/src/github.com/google/cadvisor/cmd/internal/api/handler.go:121 +0x766
  github.com/google/cadvisor/cmd/internal/api.RegisterHandlers.func1()
      /go/src/github.com/google/cadvisor/cmd/internal/api/handler.go:51 +0x75
  net/http.HandlerFunc.ServeHTTP()
      /usr/local/go/src/net/http/server.go:2007 +0x51
  net/http.(*ServeMux).ServeHTTP()
      /usr/local/go/src/net/http/server.go:2387 +0x288
  net/http.(*ServeMux).ServeHTTP()
      /usr/local/go/src/net/http/server.go:2387 +0x288
  net/http.serverHandler.ServeHTTP()
      /usr/local/go/src/net/http/server.go:2802 +0xce
  net/http.(*conn).serve()
      /usr/local/go/src/net/http/server.go:1890 +0x837

Previous write at 0x00c0004d2db8 by goroutine 61:
  github.com/google/cadvisor/perf.(*collector).UpdateStats()
      /go/src/github.com/google/cadvisor/perf/collector_libpfm.go:102 +0x95a
  github.com/google/cadvisor/manager.(*containerData).updateStats()
      /go/src/github.com/google/cadvisor/manager/container.go:658 +0x514
  github.com/google/cadvisor/manager.(*containerData).housekeepingTick()
      /go/src/github.com/google/cadvisor/manager/container.go:535 +0x23a
  github.com/google/cadvisor/manager.(*containerData).housekeeping()
      /go/src/github.com/google/cadvisor/manager/container.go:483 +0x2ee

Goroutine 50 (running) created at:
  net/http.(*Server).Serve()
      /usr/local/go/src/net/http/server.go:2928 +0x5b5
  net/http.(*Server).ListenAndServe()
      /usr/local/go/src/net/http/server.go:2825 +0x102
  main.main()
      /usr/local/go/src/net/http/server.go:3081 +0x10e9

Goroutine 61 (running) created at:
  github.com/google/cadvisor/manager.(*containerData).Start()
      /go/src/github.com/google/cadvisor/manager/container.go:108 +0x4c
  github.com/google/cadvisor/manager.(*manager).createContainerLocked()
      /go/src/github.com/google/cadvisor/manager/manager.go:1030 +0xa4e
  github.com/google/cadvisor/manager.(*manager).createContainer()
      /go/src/github.com/google/cadvisor/manager/manager.go:926 +0xbd
  github.com/google/cadvisor/manager.(*manager).watchForNewContainers.func1()
      /go/src/github.com/google/cadvisor/manager/manager.go:1182 +0x41a
==================

dashpole

I know it is still WIP, but I have a couple high-level questions.

dashpole · 2020-04-16T18:38:52Z

info/v1/container.go

@@ -899,6 +899,9 @@ type ContainerStats struct {

 	// Statistics originating from perf events
 	PerfStats []PerfStat `json:"perf_stats,omitempty"`
+
+	// Working set size
+	Wss uint64 `json:"wss,omitempty"`


Should we nest this under MemoryStats?

All memory stats come from memory cgroup so I think that wss does not fit to them.

cmd/cadvisor.go

docs/storage/prometheus.md

manager/manager.go

katarzyna-z · 2020-04-20T13:55:53Z

@dashpole It's now ready for review.

cmd/cadvisor.go

dashpole · 2020-04-21T16:35:11Z

cmd/cadvisor.go

@@ -87,6 +87,7 @@ var (
 		container.ProcessSchedulerMetrics:        struct{}{},
 		container.ProcessMetrics:                 struct{}{},
 		container.HugetlbUsageMetrics:            struct{}{},
+		container.ReferencedMetric:               struct{}{},


WDYT about ReferencedMemoryMetrics instead? Just Referenced seems a little generic...

and similarly referenced_memory for the flag string

dashpole · 2020-04-21T16:38:02Z

container/libcontainer/handler.go

+		} else {
+			stats.Referenced, err = referencedBytesStat(pids, h.cycles, *referencedResetInterval)
+			if err != nil {
+				klog.V(4).Infof("Unable to get working set size: %v", err)


s/working set size/referenced memory

dashpole · 2020-04-21T16:38:57Z

container/libcontainer/handler.go

+	return referencedKBytes * 1024, nil
+}
+
+func getReferencedBytes(pids []int) (uint64, error) {


getReferencedKBytes?

dashpole · 2020-04-21T16:41:10Z

container/libcontainer/handler.go

+		smapsFilePath := fmt.Sprintf(smapsFilePathPattern, pid)
+		smapsContent, err := ioutil.ReadFile(smapsFilePath)
+		if err != nil {
+			klog.V(3).Infof("Cannot read %s file, err: %s", smapsFilePath, err)


For logging that is expected in some scenarios, lets lower to V(5)

dashpole · 2020-04-21T16:44:34Z

container/libcontainer/handler.go

+	return referencedKBytes, nil
+}
+
+func clearReferencedBytes(pids []int, cycles uint64, resetInterval uint64) error {


Just to make sure I understand, if clearReferencedBytes errors, we will wait another resetInterval cycles before attempting to reset again, right? If so, is that better than always trying to clear after an error?

You're right, if clearReferencedBytes return error, we will wait another resetInterval but error occurs only if there is a serious issue in system (problem with writing into existing file or problem with closing previously opened file). I don't think that is a good idea to force clearing in case of errors.
User can easily observe if referenced bytes were cleared or not, seeing value of referenced bytes.

Since your last review I've introduced an option to switch off clearing of referenced bytes (setting referenced_reset_interval to 0). It's documented here. During experiments it's more user friendly to set 0 than very long reset interval if it's desired to observe referenced bytes in longer period.

see: https://github.com/brendangregg/wss#wsspl-referenced-page-flag Signed-off-by: Katarzyna Kujawa <katarzyna.kujawa@intel.com>

dashpole

lgtm

google#2495 introduced parsing smaps memory metrics, but only exposes referenced memory. As proposed in google#2634, also other metrics in here are of interest, specifically LazyFree; but generally also the others seem of interest. This commit replaces the referenced memory metrics by a generic smaps series of metrics. Runtime costs in cadvisor are not really impacted, as all the smaps parsing already happens anyway. The cardinality per container extends from 1 to 19 metrics, though. I think this is acceptable: also referenced memory scraping was disabled by default, so is scraping smaps. Resolves google#2634. Signed-off-by: Jens Erat <jens.erat@daimler.com>

katarzyna-z force-pushed the kk-wss-metric branch from 31a447c to 92b3dfc Compare April 16, 2020 07:56

katarzyna-z changed the title ~~Add working set size metric~~ WIP: Add working set size metric Apr 16, 2020

dashpole reviewed Apr 16, 2020

View reviewed changes

katarzyna-z force-pushed the kk-wss-metric branch 3 times, most recently from 7a7d5bb to f25f2fd Compare April 20, 2020 13:32

katarzyna-z changed the title ~~WIP: Add working set size metric~~ Add working set size metric Apr 20, 2020

dashpole reviewed Apr 21, 2020

View reviewed changes

katarzyna-z force-pushed the kk-wss-metric branch 3 times, most recently from 82dc9bd to cc869de Compare April 22, 2020 11:58

Add referenced bytes metric for containers

7ab5e27

see: https://github.com/brendangregg/wss#wsspl-referenced-page-flag Signed-off-by: Katarzyna Kujawa <katarzyna.kujawa@intel.com>

katarzyna-z force-pushed the kk-wss-metric branch from cc869de to 7ab5e27 Compare April 22, 2020 12:26

dashpole approved these changes Apr 22, 2020

View reviewed changes

dashpole merged commit a7d2254 into google:master Apr 22, 2020

dashpole changed the title ~~Add working set size metric~~ Add referenced memory metric May 29, 2020

edrevo mentioned this pull request Nov 3, 2020

kubelet counts active page cache against memory.available (maybe it shouldn't?) kubernetes/kubernetes#43916

Open

JensErat mentioned this pull request Dec 17, 2020

Export entire smaps memory metrics instead of only referenced_memory #2767

Closed

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add referenced memory metric #2495

Add referenced memory metric #2495

katarzyna-z commented Apr 16, 2020

katarzyna-z commented Apr 16, 2020

dashpole left a comment

dashpole Apr 16, 2020

katarzyna-z Apr 17, 2020

katarzyna-z commented Apr 20, 2020

dashpole Apr 21, 2020

dashpole Apr 21, 2020

dashpole Apr 21, 2020

dashpole Apr 21, 2020

dashpole Apr 21, 2020

dashpole Apr 21, 2020

katarzyna-z Apr 22, 2020

dashpole left a comment

Add referenced memory metric #2495

Add referenced memory metric #2495

Conversation

katarzyna-z commented Apr 16, 2020

katarzyna-z commented Apr 16, 2020

dashpole left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

katarzyna-z commented Apr 20, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dashpole left a comment

Choose a reason for hiding this comment