Currently ray start ignores --memory settings, treating them like a burstable memory request, rather than a hard limit. Are there any known workarounds to cap memory usage of Ray Core servers?
More info
Currently parameters used to control all other resources [1] act as hard limits, but memory argument (of ray.init()) and --memory switch (of ray start) are definitely not hard limits. This becomes apparent when a container with Ray server is run separately from Jupyter Notebook server with the python client app, which allows us to distinguish between the client and server-side memory usage.
[1] i.e. object store memory and CPU / GPU numbers (provisioned with object_store_memory / --object-store-memory, num_cpus / --num-cpus and num_gpus / --num-gpus).
Unfortunately there is no way to do this at the moment, but it is an active area of development.
In the meantime, the best way to enforce a hard cap is to use a container with a memory limit, but note that this will kill the Ray node if it reaches the hard cap.
If the memory usage is mainly coming from the Ray application, the best way to work around this at the moment is to request more CPUs per task, to ensure that fewer of them run in parallel. If you're seeing unusually high memory usage from Ray system-level processes like the GCS or the raylet, then it may be a bug in Ray core.
Related
I have a high-performance software server application that is expected to get increased traffic in the next few months.
I was wondering what approach or methodology is good to use in order to gauge if the server still has the capacity to handle this increased load?
I think you're looking for Stress Testing and the scenario would be something like:
Create a load test simulating current real application usage
Start with current number of users and gradually increase the load until
you reach the "increased traffic" amount
or errors start occurring
or you start observing performance degradation
whatever comes the first
Depending on the outcome you either can state that your server can handle the increased load without any issues or you will come up with the saturation point and the first bottleneck
You might also want to execute a Soak Test - leave the system under high prolonged load for several hours or days, this way you can detect memory leaks or other capacity problems.
More information: Why ‘Normal’ Load Testing Isn’t Enough
Test the product with one-tenth the data and traffic. Be sure the activity is 'realistic'.
Then consider what will happen as traffic grows -- with the RAM, disk, cpu, network, etc, grow linearly or not?
While you are doing that, look for "hot spots". Optimize them.
Will you be using web pages? Databases? Etc. Each of these things scales differently. (In other words, you have not provided enough details in your question.)
Most canned benchmarks focus on one small aspect of computing; applying the results to a specific application is iffy.
I would start by collecting base line data on critical resources - typically, CPU, memory usage, disk usage, network usage - and track them over time. If any of those resources show regular spikes where they remain at 100% capacity for more than a fraction of a second, under current usage, you have a bottleneck somewhere. In this case, you cannot accept additional load without likely outages.
Next, I'd start figuring out what the bottleneck resource for your application is - it varies between applications, but in most cases it's the bottleneck resource that stops you from scaling further. Your CPU might be almost idle, but you're thrashing the disk I/O, for instance. That's a tricky process - load and stress testing are the way to go.
If you can resolve the bottleneck by buying better hardware, do so - it's much cheaper than rewriting the software. If you can't buy better hardware, look at load balancing. If you can't load balance, you've got to look at application architecture and implementation and see if there are ways to move the bottleneck.
It's quite typical for the bottleneck to move from one resource to the next - you've got CPU to behave, but now when you increase traffic, you're spiking disk I/O; once you resolve that, you may get another CPU challenge.
We currently have a ArangoDB cluster running on version 3.0.10 for a POC with about 81 GB stored on disk and Main memory consumption of about 98 GB distributed across 5 Primary DB servers. There are about 200 million vertices and 350 million Edges. There are 3 Edge collections and 3 document collections, most of the memory(80%) is consumed due to the presence of the edges
I'm exploring methods to decrease the main memory consumption. I'm wondering if there are any methods to compress/serialize the data so that less amount of main memory is utilized.
The reason for decreasing memory is to reduce infrastructure costs, I'm willing to trade-off on speed for my use case.
Please can you let me know, if there any Methods to reduce main memory consumption for ArangoDB
It took us a while to find out that our original recommendation to set vm.overcommit_memory to 2 this is not good in all situations.
It seems that there is an issue with the bundled jemalloc memory allocator in ArangoDB with some environments.
With an vm.overcommit_memory kernel settings value of 2, the allocator had a problem of splitting existing memory mappings, which made the number of memory mappings of an arangod process grow over time. This could have led to the kernel refusing to hand out more memory to the arangod process, even if physical memory was still available. The kernel will only grant up to vm.max_map_count memory mappings to each process, which defaults to 65530 on many Linux environments.
Another issue when running jemalloc with vm.overcommit_memory set to 2 is that for some workloads the amount of memory that the Linux kernel tracks as "committed memory" also grows over time and does not decrease. So eventually the ArangoDB daemon process (arangod) may not get any more memory simply because it reaches the configured overcommit limit (physical RAM * overcommit_ratio + swap space).
So the solution here is to modify the value of vm.overcommit_memory from 2 to either 1 or 0. This will fix both of these problems.
We are still observing ever-increasing virtual memory consumption when using jemalloc with any overcommit setting, but in practice this should not cause problems.
So when adjusting the value of vm.overcommit_memory from 2 to either 0 or 1 (0 is the Linux kernel default btw.) this should improve the situation.
Another way to address the problem, which however requires compilation of ArangoDB from source, is to compile a build without jemalloc (-DUSE_JEMALLOC=Off when cmaking). I am just listing this as an alternative here for completeness. With the system's libc allocator you should see quite stable memory usage. We also tried another allocator, precisely the one from libmusl, and this also shows quite stable memory usage over time. The main problem here which makes exchanging the allocator a non-trivial issue is that jemalloc has very nice performance characteristics otherwise.
(quoting Jan Steemann as can be found on github)
Several new additions to the rocksdb storage engine were made meanwhile. We demonstrate how memory management works in rocksdb.
Many Options of the rocksdb storage engine are exposed to the outside via options.
Discussions and a research of a user led to two more options to be exposed for configuration with ArangoDB 3.7:
--rocksdb.cache-index-and-filter-blocks-with-high-priority
--rocksdb.pin-l0-filter-and-index-blocks-in-cache
--rocksdb.pin-top-level-index-and-filter
Lots of personal experience, anecdotal evidence, and some rudimentary analysis suggests that a Java server (running, typically, Oracle's 1.6 JVM) has faster response times when it's under a decent amount of load (only up to a point, obviously).
I don't think this is purely hotspot, since response times slow down a bit again when the traffic dies down.
In a number of cases we can demonstrate this by averaging response times from server logs ... in some cases it's as high as 20% faster, on average, and with a smaller standard deviation.
Can anyone explain why this is so? Is it likely a genuine effect, or are the averages simply misleading? I've seen this for years now, through several jobs, and tend to state it as a fact, but have no explanation for why.
Thanks,
Eric
EDIT a fairly large edit for wording and adding more detail throughout.
A few thoughts:
Hotspot kicks in when a piece of code is being executed significantly more than other pieces (it's the hot spot of the program). This makes that piece of code significantly faster (for the normal path) from that point forward. The rate of call after the hotspot compilation is not important, so I don't think this is causing the effect you are mentioning.
Is the effect real? It's very easy to trick yourself with statistics. Not saying you are, but be sure that all your runs are included in the result, and that all other effects (such as other programs, activity, and your monitoring program are the same in all cases. I have more than one had my monitoring program, such as top, cause a difference in behaviour). On one occasion, the performance of the application went up appreciably when the caches warmed up on the database - there was memory pressure from other applications on the same DB instance.
The Operating System and/or CPU may well be involved. The OS and CPU both actively and passively do things to improve the responsiveness of the main program as it moves from being mainly running to being mainly waiting for I/O and vice versa, including:
OS paging memory to disk while it's not being used, and back to RAM when the program is running
OS will cache frequently used disk blocks, which again may improve the application performance
CPU instruction and memory caches fill with the active program's instruction and data
Java applications particularly sensitive to memory paging effects because:
A typical Java application server will pre-allocate almost all free memory to Java. The large memory makes the application inherently more sensitive to memory effects
The generational garbage collector used to manage Java memory ends up creating new objects over a lot of pages, so each request to the application will need more page requests than in other languages. (this is true principally for 'new' objects that have not been through many garbage collections. Objects promoted to the permanent generation are actually very compactly stored)
As most available physical memory is allocated on the system, there is always a pressure on memory, and the largest, least recently run application is a perfect candidate to be pages out.
With these considerations, there is much more probability that there will be page misses and therefore a performance hit than environments with smaller memory requirements. These will be particularly manifest after Java has been idle for some time.
If you use Solaris or Mac, the excellent dTrace can trace memory and disk paging specific to an application. The JVM has numerous dTrace hooks that can be used as triggers to start and stop page monitoring.
On Solaris, you can use large memory pages (even over 1GB in size) and pin them to RAM so they will never be paged out. This should eliminate the memory page problem stated above. Remember to leave a good chunk of free memory for disk caching and for other system/maintenance/backup/management apps. I am sure that other OSes support similar features.
TL/DR: The currently running program in modern operating systems will appear to run faster after a few seconds as the OS brings the program and data pages back from disk, places frequently used disk pages in disk cache and the OS instruction and data caches will tend to be "warmer" for the main program. This effect is not unique to the JVM but is more visible due to the memory requirements of typical Java applications and the garbage collection memory model.
We've just bought a 32-core Opteron machine, and the speedups we get are a little disappointing: beyond about 24 threads we see no speedup at all (actually gets slower overall) and after about 6 threads it becomes significantly sub-linear.
Our application is very thread-friendly: our job breaks down into about 170,000 little tasks which can each be executed separately, each taking 5-10 seconds. They all read from the same memory-mapped file of size about 4Gb. They make occasional writes to it, but it might be 10,000 reads to each write - we just write a little bit of data at the end of each of the 170,000 tasks. The writes are lock-protected. Profiling shows that the locks are not a problem. The threads use a lot of JVM memory each in non-shared objects and they make very little access to shared JVM objects and of that, only a small percentage of accesses involve writes.
We're programming in Java, on Linux, with NUMA enabled. We have 128Gb RAM. We have 2 Opteron CPU's (model 6274) of 16 cores each. Each CPU has 2 NUMA nodes. The same job running on an Intel quad-core (i.e. 8 cores) scaled nearly linearly up to 8 threads.
We've tried replicating the read-only data to have one-per-thread, in the hope that most lookups can be local to a NUMA node, but we observed no speedup from this.
With 32 threads, 'top' shows the CPU's 74% "us" (user) and about 23% "id" (idle). But there are no sleeps and almost no disk i/o. With 24 threads we get 83% CPU usage. I'm not sure how to interpret 'idle' state - does this mean 'waiting for memory controller'?
We tried turning NUMA on and off (I'm referring to the Linux-level setting that requires a reboot) and saw no difference. When NUMA was enabled, 'numastat' showed only about 5% of 'allocation and access misses' (95% of cache misses were local to the NUMA node). [Edit:] But adding "-XX:+useNUMA" as a java commandline flag gave us a 10% boost.
One theory we have is that we're maxing out the memory controllers, because our application uses a lot of RAM and we think there are a lot of cache misses.
What can we do to either (a) speed up our program to approach linear scalability, or (b) diagnose what's happening?
Also: (c) how do I interpret the 'top' result - does 'idle' mean 'blocked on memory controllers'? and (d) is there any difference in the characteristics of Opteron vs Xeon's?
I also have a 32 core Opteron machine, with 8 NUMA nodes (4x6128 processors, Mangy Cours, not Bulldozer), and I have faced similar issues.
I think the answer to your problem is hinted at by the 2.3% "sys" time shown in top. In my experience, this sys time is the time the system spends in the kernel waiting for a lock. When a thread can't get a lock it then sits idle until it makes its next attempt. Both the sys and idle time are a direct result of lock contention. You say that your profiler is not showing locks to be the problem. My guess is that for some reason the code causing the lock in question is not included in the profile results.
In my case a significant cause of lock contention was not the processing I was actually doing but the work scheduler that was handing out the individual pieces of work to each thread. This code used locks to keep track of which thread was doing which piece of work. My solution to this problem was to rewrite my work scheduler avoiding mutexes, which I have read do not scale well beyond 8-12 cores, and instead use gcc builtin atomics (I program in C on Linux). Atomic operations are effectively a very fine grained lock that scales much better with high core counts. In your case if your work parcels really do take 5-10s each it seems unlikely this will be significant for you.
I also had problems with malloc, which suffers horrible lock issues in high core count situations, but I can't, off the top of my head, remember whether this also led to sys & idle figures in top, or whether it just showed up using Mike Dunlavey's debugger profiling method (How can I profile C++ code running in Linux?). I suspect it did cause sys & idle problems, but I draw the line at digging through all my old notes to find out :) I do know that I now avoid runtime mallocs as much as possible.
My best guess is that some piece of library code you are using implements locks without your knowledge, is not included in your profiling results, and is not scaling well to high core-count situations. Beware memory allocators!
I'm sure the answer will lie in a consideration of the hardware architecture. You have to think of multi core computers as if they were individual machines connected by a network. In fact that's all that Hypertransport and QPI are.
I find that to solve these scalability problems you have to stop thinking in terms of shared memory and start adopting the philosophy of Communicating Sequential Processes. It means thinking very differently, ie imagine how you would write the software if your hardware was 32 single core machines connected by a network. Modern (and ancient) CPU architectures are not designed to give unfettered scaling of the sort you're after. They are designed to allow many different processes to get on with processing their own data.
Like everything else in computing these things go in fashions. CSP dates back to the 1970s, but the very modern and Java derived Scala is a popular embodiment of the concept. See this section on Scala concurrency on Wikipedia.
What the philosophy of CSP does is force you to design a data distribution scheme that fits your data and the problem you're solving. That's not necessarily easy, but if you manage it then you have a solution that will scale very well indeed. Scala may make it easier to develop.
Personally I do everything in CSP and in C. It's allowed me to develop a signal processing application that scales perfectly linearly from 8 cores to several thousand cores (the limit being how big my room is).
The first thing you're going to have to do is actually use NUMA. It isn't a magic setting that you turn on, you have to exploit it in your software's architecture. I don't know about Java, but in C one would bind a memory allocation to a specific core's memory controller (aka memory affinity), and similarly for threads (core affinity) in cases where the OS doesn't get the hint.
I presume that your data doesn't break down into 32 neat, discrete chunks? It's difficult to give advice without knowing exactly the data flows implicit in your program. But think about it in terms of data flow. Draw it out even; Data Flow Diagrams are useful for this (another ancient graphical formal notation). If your picture shows all your data going through a single object (eg through a single memory buffer) then it's going to be slow...
I assume you have optimized your locks, and synchronization made a minimum. In such a case, it still depends a lot on what libraries you are using to program in parallel.
One issue that can happen even if you have no synchronization issue, is memory bus congestion. This is very nasty and difficult to get rid of.
All I can suggest is somehow make your tasks bigger and create fewer tasks. This depends highly on the nature of your problem. Ideally you want as many tasks as the number of cores/threads, but this is not easy (if possible) to achieve.
Something else that can help is to give more heap to your JVM. This will reduce the need to run Garbage Collector frequently, and speeds up a little.
does 'idle' mean 'blocked on memory controllers'
No. You don't see that in top. I mean if the CPU is waiting for memory access, it will be shown as busy. If you have idle periods, it is either waiting for a lock, or for IO.
I'm the Original Poster. We think we've diagnosed the issue, and it's not locks, not system calls, not memory bus congestion; we think it's level 2/3 CPU cache contention.
To reiterate, our task is embarrassingly parallel so it should scale well. However, one thread has a large amount of CPU cache it can access, but as we add more threads, the amount of CPU cache each process can access gets lower and lower (the same amount of cache divided by more processes). Some levels on some architectures are shared between cores on a die, some are even shared between dies (I think), and it may help to get "down in the weeds" with the specific machine you're using, and optimise your algorithms, but our conclusion is that there's not a lot we can do to achieve the scalability we thought we'd get.
We identified this as the cause by using 2 different algorithms. The one which accesses more level 2/3 cache scales much worse than the one which does more processing with less data. They both make frequent accesses to the main data in main memory.
If you haven't tried that yet: Look at hardware-level profilers like Oracle Studio has (for CentOS, Redhat, and Oracle Linux) or if you are stuck with Windows: Intel VTune. Then start looking at operations with suspiciously high clocks per instruction metrics. Suspiciously high mean a lot higher than the same code on a single-numa, single-L3-cache machine (like current Intel desktop CPUs).
Note: I am aware of the question Memory management in memory intensive application, however that question appears to be about applications that make frequent memory allocations, whereas my question is about applications intentionally designed to consume as much physical memory as is safe.
I have a server application that uses large amounts of memory in order to perform caching and other optimisations (think SQL Server). The application runs on a dedicated machine, and so can (and should) consume as much memory as it wants / is able to in order to speed up and increase throughput and response times without worry of impacting other applications on the system.
The trouble is that if memory usage is underestimated, or if load increases its possible to end up with nasty failures as memory allocations fail - in this situation obviously the best thing to do is to free up memory in order to prevent the failure at the expense of performance.
Some assumptions:
The application is running on a dedicated machine
The memory requirements of the application exceed the physical memory on the machine (that is, if additional memory was available to the application it would always be able to use that memory to in some way improve response times or throughput)
The memory is effectively managed in a way such that memory fragmentation is not an issue.
The application knows what memory can be safely freed, and what memory should be freed first for the least performance impact.
The app runs on a Windows machine
My question is - how should I handle memory allocations in such an application? In particular:
How can I predict whether or not a memory allocation will fail?
Should I leave a certain amount of memory free in order to ensure that core OS operations remain responsive (and don't in that way adversely impact the applications performance), and how can I find out how much memory that is?
The core objective is to prevent failures as a result of using too much memory, while at the same time using up as much memory as possible.
I'm a C# developer, however my hope is that the basic concepts for any such app are the same regardless of the language.
In linux, the memory usage percentage is divided into following levels.
0 - 30% - no swapping
30 - 60% - swap dirty pages only
60 - 90% - swap clean pages also based on LRU policy.
90% - Invoke OOM(Out of memory) killer and kill the process consuming maximum memory.
check this - http://linux-mm.org/OOM_Killer
In think windows might have similar policy, so you can check the memory stats and make sure you never get to the max threshold.
One way to stop consuming more memory is to go to sleep and give more time for memory cleanup threads to run.
That is a very good question, and bound to be subjective as well, because the very nature of the fundamental of C# is that all memory management is done by the runtime, i.e. Garbage Collector. The Garbage Collector is a non-deterministic entity that manages and sweeps the memory for reclaiming, depending on how often the memory gets fragmented, the GC will kick in hence to know in advance is not easy thing to do.
To properly manage the memory sounds tedious but common sense applies, such as the using clause to ensure an object gets disposed. You could put in a single handler to trap the OutOfMemory Exception but that is an awkward way, since if the program has run out of memory, does the program just seize up, and bomb out, or should it wait patiently for the GC to kick in, again determining that is tricky.
The load of the system can adversely affect the GC's job, almost to a point of a Denial of Service, where everything just grind to a halt, again, since the specifications of the machine, or what is the nature of that machine's job is unknown, I cannot answer it fully, but I'll assume it has loads of RAM..
In essence, while an excellent question, I think you should not worry about it and leave it to the .NET CLR to handle the memory allocation/fragmentation as it seems to do a pretty good job.
Hope this helps,
Best regards,
Tom.
Your question reminds me of an old discussion "So what's wrong with 1975 programming ?". The architect of varnish-cache argues, that instead of telling the OS to get out of the way and manage all memory yourself, you should rather cooperate with the OS and let it understand what you intend to do with memory.
For example, instead of simply reading data from disk, you should use memory-mapped files. This allows the OS to apply its LRU algorithm to write-back data to disk when memory becomes scarce. At the same time, as long as there is enough memory, your data will stay in memory. Thus, your application may potentially use all memory, without risking getting killed.