Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
See https://github.com/golang/go/issues/9849 Go has no limit, you use ulimit to control. -Original Message- >From: Kevin Chadwick >Sent: Jan 23, 2020 10:26 AM >To: golang-nuts@googlegroups.com >Subject: Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation? > >On 2020-01-23 14:18, robert engels wrote: >> There is nothing “special” about it - generally the Go process calls >> “malloc()” and fails with OOM (unable to expand the process memory size), >> but the OS can kill processes in a low system memory condition without them >> calling malloc (OOM killer kills the hogs). If you process is dying due to >> the OOM killer you have configuration problems. >> > >Because GO has it's own limit right? Whereas firefox/chrome didn't bother, >likely because who are firefox/chrome to limit your systems memory use when it >is the OS job. I guess GO limits it for panics or because Linux has a >questionable default of unlimited or maybe because it supports so many >platforms, otherwise it does not make any sense to me. The point is that if >polling has race issues then I would investigate whether an OOM return/panic >might be a timely indicator to the application/routine to clean house. > >> When malloc() fails in a GC system, it could be because the free space is >> fragmented. In a compacting and moving GC, it will shift objects around to >> make room (Go does not does this, most Java collectors do). Additionally, >> what I was primarily pointing out, rather than failing the GC will free >> “soft refs” to make room. > >Right, I believe tiny go does some compacting but you are better off without >dynamic memory at all on a micro without an MMU. This compacting doesn't really >solve the issue in a dynamic world with an MMU either, but just makes the >system >more predictable, doesn't it? > >-- >You received this message because you are subscribed to the Google Groups >"golang-nuts" group. >To unsubscribe from this group and stop receiving emails from it, send an >email to golang-nuts+unsubscr...@googlegroups.com. >To view this discussion on the web visit >https://groups.google.com/d/msgid/golang-nuts/7fea4901-16df-93e7-8a7c-9b95beb9a33e%40gmail.com. -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/58993198.1164.1579797754762%40wamui-bella.atl.sa.earthlink.net.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
On 2020-01-23 14:18, robert engels wrote: > There is nothing “special” about it - generally the Go process calls > “malloc()” and fails with OOM (unable to expand the process memory size), but > the OS can kill processes in a low system memory condition without them > calling malloc (OOM killer kills the hogs). If you process is dying due to > the OOM killer you have configuration problems. > Because GO has it's own limit right? Whereas firefox/chrome didn't bother, likely because who are firefox/chrome to limit your systems memory use when it is the OS job. I guess GO limits it for panics or because Linux has a questionable default of unlimited or maybe because it supports so many platforms, otherwise it does not make any sense to me. The point is that if polling has race issues then I would investigate whether an OOM return/panic might be a timely indicator to the application/routine to clean house. > When malloc() fails in a GC system, it could be because the free space is > fragmented. In a compacting and moving GC, it will shift objects around to > make room (Go does not does this, most Java collectors do). Additionally, > what I was primarily pointing out, rather than failing the GC will free “soft > refs” to make room. Right, I believe tiny go does some compacting but you are better off without dynamic memory at all on a micro without an MMU. This compacting doesn't really solve the issue in a dynamic world with an MMU either, but just makes the system more predictable, doesn't it? -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/7fea4901-16df-93e7-8a7c-9b95beb9a33e%40gmail.com.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
There is nothing “special” about it - generally the Go process calls “malloc()” and fails with OOM (unable to expand the process memory size), but the OS can kill processes in a low system memory condition without them calling malloc (OOM killer kills the hogs). If you process is dying due to the OOM killer you have configuration problems. When malloc() fails in a GC system, it could be because the free space is fragmented. In a compacting and moving GC, it will shift objects around to make room (Go does not does this, most Java collectors do). Additionally, what I was primarily pointing out, rather than failing the GC will free “soft refs” to make room. > On Jan 23, 2020, at 6:13 AM, Kevin Chadwick wrote: > > On 2020-01-20 18:57, Robert Engels wrote: >> This is solved pretty easily in Java using soft references and a hard memory >> cap. >> >> Similar techniques may work here. > > One of the only things I dislike about GO compared to C is the arbitrary > memory > allocation but it has great benefits in coding time and I expect you can > handle > an array allocation panic etc. and keep track of your buffers. > > I know that OpenBSD sets limits by default which need to be raised for > chrome/firefox to prevent OOM death etc. which isn't the case on Linux without > default limits (last I heard). I believe Linux kills the hogging process > instead! > > I seem to remember OpenBSD devs saying the OS provides opportunity for Firefox > to manage it's own OOM condition with these limits in place. I took that to > mean > that Linux defaults have made it difficult to handle this properly in general, > but I may lack understanding of the general issue? > > -- > You received this message because you are subscribed to the Google Groups > "golang-nuts" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to golang-nuts+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/golang-nuts/97c30a53-1b99-6424-fd00-9a7ae10a77f6%40gmail.com. -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/8F45240D-E57E-4655-98BE-4E103A222405%40ix.netcom.com.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
On 2020-01-20 18:57, Robert Engels wrote: > This is solved pretty easily in Java using soft references and a hard memory > cap. > > Similar techniques may work here. One of the only things I dislike about GO compared to C is the arbitrary memory allocation but it has great benefits in coding time and I expect you can handle an array allocation panic etc. and keep track of your buffers. I know that OpenBSD sets limits by default which need to be raised for chrome/firefox to prevent OOM death etc. which isn't the case on Linux without default limits (last I heard). I believe Linux kills the hogging process instead! I seem to remember OpenBSD devs saying the OS provides opportunity for Firefox to manage it's own OOM condition with these limits in place. I took that to mean that Linux defaults have made it difficult to handle this properly in general, but I may lack understanding of the general issue? -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/97c30a53-1b99-6424-fd00-9a7ae10a77f6%40gmail.com.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
This is actually SoftRef - which are like a WeakRef but they are only collected under memory pressure. If the WeakRef package works, I assume it could be modified to enable “soft ref” like functionality. It was my understanding that you need GC/runtime support to truly make this work, but maybe they have an Unsafe/CGO way. I haven’t really researched the WeakRef packages for Go. > On Jan 20, 2020, at 4:58 PM, Eric S. Raymond wrote: > > Robert Engels : >> This is solved pretty easily in Java using soft references and a hard memory >> cap. > > That'd be nice, but the onnly weak-references package I've found doesn't seem > to allow > more than one weakref per target. That's really annoying, because my use case > is > a target object for a many-to-one mapping that should become GCable when the > last of its source objects is GCed. > > Is there a weakrefs implementation out tere that will do that? > -- > http://www.catb.org/~esr/;>Eric S. Raymond > > > -- > You received this message because you are subscribed to the Google Groups > "golang-nuts" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to golang-nuts+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/golang-nuts/20200120225853.GA75815%40thyrsus.com. -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/A97065E1-66FD-42A5-9E6B-7AFA5CD96A68%40ix.netcom.com.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
Robert Engels : > This is solved pretty easily in Java using soft references and a hard memory > cap. That'd be nice, but the onnly weak-references package I've found doesn't seem to allow more than one weakref per target. That's really annoying, because my use case is a target object for a many-to-one mapping that should become GCable when the last of its source objects is GCed. Is there a weakrefs implementation out tere that will do that? -- http://www.catb.org/~esr/;>Eric S. Raymond -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/20200120225853.GA75815%40thyrsus.com.
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
This is solved pretty easily in Java using soft references and a hard memory cap. Similar techniques may work here. > On Jan 20, 2020, at 11:22 AM, Christian Mauduit wrote: > > Hi, > > That is a generic question and I think that if you want to keep an approach > with a "global indicator of how much memory is used", your approach is OK. > You might also want to store this information of "should I throttle" in a > cache or something, the cache could be just a shared atomic flag. > > But, from my experience, ugly OOMs can come very fast. So a ticker might just > not catch them. On modern hardware, you can allocate gigabytes of memory > within seconds, and hit the roof before your ticker based watcher notices > anything. > > The best rules I have seen to enforce proper memory control is: > > - channels are blessed as they are limited in size > - bind everything else to a limit. More specifically, enfore limits for: > - maps (limit the number of entries) > - arrays > - generally speaking, anything which can *grow* > > It is tedious, but on top of avoiding the "my ticker does not run often > enough" problem, you will also be able to take a decision (should I allocate > or not) based on local, actionable facts. The problem with the global > approach is that at some point, when the program is running out of memory, > you will not know *WHAT* is taking so much space. So you might end up > stopping allocating Foos when the problem is that there are too many Bars > allocated. I have seen this happening over and over. > > So from my experience, just: > > - observe, monitor your program to see what is taking most memory > - enforce limitation on that part to ensure it is bounded > - make this configurable as you will need to fine-tune it > - write a test that would OOM without memory control > > Sometimes the naive approach of those low-level limits does not work. > > Example: you have a map[int64]string and you can affort ~1Gb for it, but the > size of string can range from 1 byte to 1Mb, with an average of 1Kb, and you > need to store 1M keys. With such a case, if I want to be serious about OOMs > and if it is a real problem, I would bite the bullet and simply track the > total amount of bytes stored. Maybe not the precise amount of data, summing > the length of all strings could be enough. > > But the bottom line is -> I fear there is no "one size fits all" solution for > this. > > It is expected that library providers do not implement this as they wish to > offer generic multi-purpose tools, and memory enforcement is rather an > application, final-product requirement. > > Best luck, OOMs are hard. > > Christian. > >> On 20/01/2020 09:22, Urjit Singh Bhatia wrote: >> Hi folks, >> I am trying to figure out if someone has a decent solution for max memory >> usage/mem-pressure so far. I went through some of the github issues related >> to this (SetMaxHeap proposals and related discussions) but most of them are >> still under review: >> * https://go-review.googlesource.com/c/go/+/46751 >> * https://github.com/golang/go/issues/16843 >> * https://github.com/golang/go/issues/23044 >> My use case is that I have an in-memory data store and producers can >> potentially drive it to OOM, causing the server to drop all the data. I'd >> like to avoid it as much as possible by potentially informing the producers >> to slow down/refuse new jobs for a bit till some data is drained if I has >> some sort of a warning system. I think rabbitmq does something similar >> https://www.rabbitmq.com/memory.html >> * How are others in the community handling these situations? It seems >>like most of database-ish implementations will just OOM and die. >> * Has anyone implemented a mem pressure warning mechanism? >> Does something like this make sense for detecting high mem usage? (Of course >> other programs can randomly ask for memory from the OS so this isn't going >> to be leak-proof) >>funcmemWatcher(highWatermarkuint64, memCheckIntervaltime.Duration) { >>s:={} >>// Count how many times we are consecutively beyond a mem watermark >>highWatermarkCounter:=-1 >>forceGCInterval:=10 >>forrangetime.NewTicker(memCheckInterval).C { >>runtime.ReadMemStats(s) >>ifs.NextGC >=highWatermark { >>log.Println("Approaching highWatermark") // Approaching highWatermark >>continue >>} >>// Crossing highWatermark >>ifs.HeapAlloc >=highWatermark { >>ifhighWatermarkCounter <0{ >>// Transitioning beyond highWatermark >>log.Println("High mem usage detected! Crossed high watermark") >>// Start counters >>highWatermarkCounter=0 >>forceGCInterval=10 >>log.Println("Reject/Throttle new work till usage reduces") >>continue >>} else{ >>// Still above highWatermark >>highWatermarkCounter++ >>ifhighWatermarkCounter >=forceGCInterval { >>log.Println("Forcing GC...") >>runtime.GC() >>
Re: [go-nuts] Is there some kind of a MaxHeapAlarm implementation?
Hi, That is a generic question and I think that if you want to keep an approach with a "global indicator of how much memory is used", your approach is OK. You might also want to store this information of "should I throttle" in a cache or something, the cache could be just a shared atomic flag. But, from my experience, ugly OOMs can come very fast. So a ticker might just not catch them. On modern hardware, you can allocate gigabytes of memory within seconds, and hit the roof before your ticker based watcher notices anything. The best rules I have seen to enforce proper memory control is: - channels are blessed as they are limited in size - bind everything else to a limit. More specifically, enfore limits for: - maps (limit the number of entries) - arrays - generally speaking, anything which can *grow* It is tedious, but on top of avoiding the "my ticker does not run often enough" problem, you will also be able to take a decision (should I allocate or not) based on local, actionable facts. The problem with the global approach is that at some point, when the program is running out of memory, you will not know *WHAT* is taking so much space. So you might end up stopping allocating Foos when the problem is that there are too many Bars allocated. I have seen this happening over and over. So from my experience, just: - observe, monitor your program to see what is taking most memory - enforce limitation on that part to ensure it is bounded - make this configurable as you will need to fine-tune it - write a test that would OOM without memory control Sometimes the naive approach of those low-level limits does not work. Example: you have a map[int64]string and you can affort ~1Gb for it, but the size of string can range from 1 byte to 1Mb, with an average of 1Kb, and you need to store 1M keys. With such a case, if I want to be serious about OOMs and if it is a real problem, I would bite the bullet and simply track the total amount of bytes stored. Maybe not the precise amount of data, summing the length of all strings could be enough. But the bottom line is -> I fear there is no "one size fits all" solution for this. It is expected that library providers do not implement this as they wish to offer generic multi-purpose tools, and memory enforcement is rather an application, final-product requirement. Best luck, OOMs are hard. Christian. On 20/01/2020 09:22, Urjit Singh Bhatia wrote: Hi folks, I am trying to figure out if someone has a decent solution for max memory usage/mem-pressure so far. I went through some of the github issues related to this (SetMaxHeap proposals and related discussions) but most of them are still under review: * https://go-review.googlesource.com/c/go/+/46751 * https://github.com/golang/go/issues/16843 * https://github.com/golang/go/issues/23044 My use case is that I have an in-memory data store and producers can potentially drive it to OOM, causing the server to drop all the data. I'd like to avoid it as much as possible by potentially informing the producers to slow down/refuse new jobs for a bit till some data is drained if I has some sort of a warning system. I think rabbitmq does something similar https://www.rabbitmq.com/memory.html * How are others in the community handling these situations? It seems like most of database-ish implementations will just OOM and die. * Has anyone implemented a mem pressure warning mechanism? Does something like this make sense for detecting high mem usage? (Of course other programs can randomly ask for memory from the OS so this isn't going to be leak-proof) funcmemWatcher(highWatermarkuint64, memCheckIntervaltime.Duration) { s:={} // Count how many times we are consecutively beyond a mem watermark highWatermarkCounter:=-1 forceGCInterval:=10 forrangetime.NewTicker(memCheckInterval).C { runtime.ReadMemStats(s) ifs.NextGC >=highWatermark { log.Println("Approaching highWatermark") // Approaching highWatermark continue } // Crossing highWatermark ifs.HeapAlloc >=highWatermark { ifhighWatermarkCounter <0{ // Transitioning beyond highWatermark log.Println("High mem usage detected! Crossed high watermark") // Start counters highWatermarkCounter=0 forceGCInterval=10 log.Println("Reject/Throttle new work till usage reduces") continue } else{ // Still above highWatermark highWatermarkCounter++ ifhighWatermarkCounter >=forceGCInterval { log.Println("Forcing GC...") runtime.GC() forceGCInterval=forceGCInterval *2 // Some kind of back-off } } } else{ ifhighWatermarkCounter >=0{ // reset counters - back under the highWatermark log.Warn().Msg("Mem usage is back under highWatermark") highWatermarkCounter=-1 forceGCInterval=10 } } } } -- You received this message because you are subscribed to the Google Groups "golang-nuts" group.
[go-nuts] Is there some kind of a MaxHeapAlarm implementation?
Hi folks, I am trying to figure out if someone has a decent solution for max memory usage/mem-pressure so far. I went through some of the github issues related to this (SetMaxHeap proposals and related discussions) but most of them are still under review: - https://go-review.googlesource.com/c/go/+/46751 - https://github.com/golang/go/issues/16843 - https://github.com/golang/go/issues/23044 My use case is that I have an in-memory data store and producers can potentially drive it to OOM, causing the server to drop all the data. I'd like to avoid it as much as possible by potentially informing the producers to slow down/refuse new jobs for a bit till some data is drained if I has some sort of a warning system. I think rabbitmq does something similar https://www.rabbitmq.com/memory.html - How are others in the community handling these situations? It seems like most of database-ish implementations will just OOM and die. - Has anyone implemented a mem pressure warning mechanism? Does something like this make sense for detecting high mem usage? (Of course other programs can randomly ask for memory from the OS so this isn't going to be leak-proof) func memWatcher(highWatermark uint64, memCheckInterval time.Duration) { > s := {} > // Count how many times we are consecutively beyond a mem watermark > highWatermarkCounter := -1 > forceGCInterval := 10 > for range time.NewTicker(memCheckInterval).C { > runtime.ReadMemStats(s) > if s.NextGC >= highWatermark { > log.Println("Approaching highWatermark") // Approaching highWatermark > continue > } > // Crossing highWatermark > if s.HeapAlloc >= highWatermark { > if highWatermarkCounter < 0 { > // Transitioning beyond highWatermark > log.Println("High mem usage detected! Crossed high watermark") > // Start counters > highWatermarkCounter = 0 > forceGCInterval = 10 > log.Println("Reject/Throttle new work till usage reduces") > continue > } else { > // Still above highWatermark > highWatermarkCounter++ > if highWatermarkCounter >= forceGCInterval { > log.Println("Forcing GC...") > runtime.GC() > forceGCInterval = forceGCInterval * 2 // Some kind of back-off > } > } > } else { > if highWatermarkCounter >= 0 { > // reset counters - back under the highWatermark > log.Warn().Msg("Mem usage is back under highWatermark") > highWatermarkCounter = -1 > forceGCInterval = 10 > } > } > } > } > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/3408da0e-f420-4b3d-98f0-d0a06e95d398%40googlegroups.com.