Go, Containers, and the Linux Scheduler
riverphillips.dev
riverphillips.dev
I wondered for a while if docker could make a fake /proc/cpuinfo that apps could parse that just listed "docker cpus" allocated to the job, but upon further reflection, that probably wouldn't work for many reasons.
(There are exceptions, such as documented here: https://kubernetes.io/docs/tasks/administer-cluster/cpu-mana...)
flask/gunicorn will fork 16 processes (by reading /proc/cpuinfo and counting cores) all of which will try to share 0.5 cores worth of CPU power (maybe spread over many physical CPUs; I don't really care about that).
I can solve this by passing a flag to my application; my complaint is more that apps shouldn't consult /proc/cpuinfo, but have another standard interface to ask "what should I set my max parallelism (NOT CONCURRENCY, ROB) so my worker threads get adequate CPU time so the framework doesn't time out on startup.
Been a bit but I do believe that dotnet does this exact behavior. Sounds like gunicorn needs a pr to mimic, if they want to replicate this.
Caveats: up to the number of cores exposed to your VM. I also believe the later versions of K8s let you do some degree of workload-core pinning and I don’t yet know how that interacts with core availability .
If you set cpus to 2 and another process sets to 1 and you both try to use all CPUs all out, you’ll get 66% and they’ll get 33%.
This isn’t the same as cpusets, which work differently.
That's not true at all. Take a look at `cpu.cfs_quota_us` in https://kernel.googlesource.com/pub/scm/linux/kernel/git/glo...
It's a hard time limit. It doesn't care about contention at all.
`cpu.shares` is relative, for choosing which process gets scheduled, and how often, but the CFS quota is a hard limit on runtime.
Not at my work environment right now, don't remember the parameters I used.
docker run --cpus 0.1 --rm -it progrium/stress --cpu 16
The true objective function I'm trying to optimize isn't just "save money" or "don't waste CPU resources", but rather "get a million different workloads to run smoothly on a large collection of resources, ensuring that revenue-critical jobs always can run, while any spare capacity is available for experimenters, up to some predefined limits determined by power capacity, staying within the overall budget, and not pissing off any really powerful users." (well, that's really just a simplified approximation)
At work at my new job we run mixed workloads in k8s and I don't really see a problem, but we also don't instrument well enough that I could say for sure. In our case it usually just makes sense to not oversubscribe machines (Google oversubscribed and then paid a cost due to preemptions and random job failures that got masked over by retries) by getting more machines.
I know the JVM automatically uses it, and there's a popular library for Go that sets GONAXPROCS using it.
Just setting it to 16 will satisfy 99% of users.
Fundamentally, intelligent apps need to interrogate their environment to make concurrency decisions. But I agree- Go would probably work best if it just picked a standard parallelism constant like 16 and just let users know that can be tuned if they have additional context.
Most people are running on systems where the CPU capacity varies and they haven't even noticed. For example in EC2 there are 8 victim CPUs that handle all the network interrupts, so if you have an instance type with 32 CPUs, you already have 24 that are faster than the others. Practically nobody even notices this effect.
Fascinating. Could you share any (all) more detail on this that you know? Is it a specific instance type, only ones that use nitro? (or only ones without?) This might be related to a problem I've seen in the wild but never tracked down...
This is new to me. What is this… behavior? What keywords should I use to find any details about it?
The only thing that rings a bell is requests/limit parameters of a pod but you can't change them on an existing pod AFAIK.
What I tried: I created a "Burstable" Pod and run `nproc` [0] on it. It returned N CPUs (N > 1).
Then I created a "Guaranteed QoS" Pod with both requests and limit set to 1 CPU. `nproc` returned N CPUs on it.
I went back to the "Burstable" Pod. It returned N.
I created a fresh "Burstable" Pod and run `nproc` on it, got N again. Please note that the "Guaranteed QoS" Pod is still running.
> Pods with Guaranteed QoS will have exactly the number of CPUs they asked for, no more or less
Well, in my case I asked for 1 CPU and got more, i.e. N CPUs.
Also, please note that Pods might ask for fractional CPUs.
[0]: coreutils `nproc` program uses `sched_getaffinity` syscall under the hood, at least on my system. I've just checked it with `strace` to be sure.
FWIW, `taskset` uses the same syscall as `nproc` (according to `strace`).
# taskset -c -p 1
pid 1's current affinity list: 1
# nproc
1
In my case it's `"cpuManagerPolicy": "none"` and I suppose you're using `"static"` policy.
Well, TIL. Thanks!
> Print the number of processing units available to the current process, which may be less than the number of online processors. If this information is not accessible, then print the number of processors installed
https://www.gnu.org/software/coreutils/manual/html_node/npro...
https://www.flamingspork.com/blog/2020/11/25/why-you-should-...
if [[ -e /sys/fs/cgroup/cpu/cpu.cfs_quota_us ]] && [[ -e /sys/fs/cgroup/cpu/cpu.cfs_period_us ]]; then
GOMAXPROCS=$(perl -e 'use POSIX; printf "%d\n", ceil($ARGV[0] / $ARGV[1])' "$(cat /sys/fs/cgroup/cpu/cpu.cfs_quota_us)" "$(cat /sys/fs/cgroup/cpu/cpu.cfs_period_us)")
else
GOMAXPROCS=$(nproc)
fi
export GOMAXPROCS
Or use https://github.com/uber-go/automaxprocs, which is very comprehensive, but is a bunch of code for what should be a simple task.
Oh they did, they're a modern day IBM.
> Containers are a crappy abstraction
They're one of the best abstractions we have (so far) because they contain only the application and what it needs.
Delusion level: over 9900.
I'm yet to find a container that contains only the application and what it needs. Most of the time I find that they contain at least libc and libpthreads (which are already present on the host, so not needed). More often I find metric tonnes of garbage that was not necessary by any metric, but was just too hard to remove, so was allowed to stay.
This exists: https://github.com/lxc/lxcfs
lxcfs is a FUSE filesystem that mocks /proc by inferring cgroup values in a way that makes other applications and libraries work without having to care about whether it runs in a container (to the best of its ability - there are definitely caveats).
One such example is that /proc/uptime should reflect the uptime of the container, not the host; additionally /proc/cpuinfo reflects the number of CPUs as a combination of cpu.max and cpuset.cpus (whichever the lower bound is).
As others also mentioned, inferring the number of CPUs could also be done using the sched_getaffinity syscall - this doesn't depend on /proc/cpuinfo, so depending on the library you're using you might be in a pickle.
From Linux docs:
- cpu.shares: The weight of each group living in the same hierarchy, that
translates into the amount of CPU it is expected to get. Upon cgroup creation,
each group gets assigned a default of 1024. The percentage of CPU assigned to
the cgroup is the value of shares divided by the sum of all shares in all
cgroups in the same level.
- cpu.cfs_period_us: The duration in microseconds of each scheduler period, for
bandwidth decisions. This defaults to 100000us or 100ms. Larger periods will
improve throughput at the expense of latency, since the scheduler will be able
to sustain a cpu-bound workload for longer. The opposite of true for smaller
periods. Note that this only affects non-RT tasks that are scheduled by the
CFS scheduler.
- cpu.cfs_quota_us: The maximum time in microseconds during each cfs_period_us
in for the current group will be allowed to run. For instance, if it is set to
half of cpu_period_us, the cgroup will only be able to peak run for 50 % of
the time. One should note that this represents aggregate time over all CPUs
in the system. Therefore, in order to allow full usage of two CPUs, for
instance, one should set this value to twice the value of cfs_period_us.If you set request / limit to 1 core but you run on 64 cores node , then you runtime will see that which will bring performance down.
I still believe it's best to set these variables regardless of cpu limits and/or cpu shares
Other people have encountered it too hence libraries like Automaxprocs existing and issues being open with Go for it.
This aligns with what others have reported on the Go runtime issue open for this.
"When go.uber.org/automaxprocs rolled out at Uber, the effect on containerized Go services was universally positive. At least at the time, CFS imposed such heavy penalties on Go binaries exceeding their CPU allotment that properly tuning GOMAXPROCS was a significant latency and throughput improvement."
https://github.com/golang/go/issues/33803#issuecomment-14308...
I'll try and clarify this. I think this is how the sympton presents but I should be clearer.
One caveat is that an application can detect when --cpu is used as I think it's using cpuset. When quota are used it cannot detect and more threads than necessary will likely be spawned
This is rather strong language without any real qualifiers. It is definitely not "mostly useless". Shares and quotas are for different use-cases, that's all. Understand your use-case and choose accordingly.
imagine some non critical system like blog-service suddenly causing 2-3% new order creation failures
When you want to limit the max CPU time available to a container use quotas (--cpus). When you want to set relative priorities (compared to other containers/processes), use shares.
These two options can be combined, it all depends on what you need.
CPU reservations are limits, just implicit ones and declared as guarantees.
So let the Go runtime use all the CPUs available, and let the Linux scheduler throttle according to your declared reservations if the CPU is contended for.
You also cannot achieve a QoS class of Guaranteed without both CPU and Memory limits, so the pod might be evicted at some point.
> You also cannot achieve a QoS class of Guaranteed without both CPU and Memory limits, so the pod might be evicted at some point.
Evicted due to node pressure - yes (but if all other pods also don't have limits it doesn't matter). For preemption QoS is not factored in the decision [0]
[0] - https://kubernetes.io/docs/concepts/scheduling-eviction/pod-...
I would be curious to see this discussed but your article only states that people think you need limit to ensure CPU for all pods.
As the node fills up with more and more other pods, it's possible that a pod that was running just fine a moment ago is crawling to a halt.
Limits allow me to simulate the same behavior and plan for it by doing the right capacity planning.
They are not the only way to approach it! But they are the simplest way to so it.
You need to do it anyway because a service might progressively need more resources as it's getting more traffic, even if you're not adding any other pod.
But having _neighbours_ affecting the behaviour of your workload is precisely what creates the kind of fatigue that then results in people claiming that it's hard to run k8s workloads. K8s is highly dynamical, pods can get scheduled on a node by chance sometimes and on some clusters; pagers will ring, incidents will be created for conditions that may solve themselves because of another deployment (possibly of another team) happening.
Overcommit/bursting is an advanced cost saving feature.
Let me say it again: splitting up a large machine into smaller parts that can use the unused capacity of other parts in order to reduce waste is an advanced feature!
The problem is that the request/limits feature is presented in the configuration spec and in the documentation in a deceptively simple way and we're tricked to think it's a basic feature.
Not all companies have ops teams that are well equipped to do more sophisticated things. My advice for those teams who cannot setup full automation around capacity management is to just not use this advanced features.
An alternative is to just use smaller dedicated nodes and (anti)affinity rules, so you always understand which pods go with which other pods. It's clunky but it's actually easier to reason about what's going to happen.
EDIT: typos
That's an interesting take. For those of us who once used microcomputers back in the 1980s like Commodore 64s, Apple IIs, and IBM PCs running MS-DOS, sure. But time-sharing systems have been around since the 1960s and Unix dates back to the 1970s, and multi-user logins and resource controls like ulimits have been part of the picture for a very long time.
We've had plenty of time to get used to multitasking and resource contention and management. Is it complicated? It can be. But if you consider that containers are basically a program lifecycle/deployment and resource control abstraction on an OS that's had these features (maybe not cgroups and namespaces, but similar ones) since its birth, it's not really all that advanced.
When you're not latency sensitive but aiming at optimizing throughput you can achieve a pretty good utilization of the underlying resource. This is what resource<limit setting in k8s is good for: batch workload.
Batch workload still exists and it's important but the proportion of batch vs latency sensitive workload shifted significantly from the 70s to internet age. Request handlers that power that sit in the critical path for the user experience of a web or mobile app are not only latency sensitive but also _tail_ latency sensitive.
Tail latency often poses significant problems because stragglers can often determine the total duration of the interaction as perceived by the user, even if only one straggler request suffers from slowdowns.
While there are tricks to deal with tail latency while also improving utilization (i.e. reducing waste) they are hard to implement in practice and/or don't generalize well across workloads.
One thing always works and is your best option as a starting point: stop worrying about idle CPU in an _average_ usage metric. The behaviour during CPU usage spikes (even spikes that are shorter than your scraping period of your favourite monitoring tool!) determine latency envelope of your service. Measure how much CPU you need during those and plan accordingly.
Ideally it should be possible to some low priority (batch) process use that idle CPU, but AFAIK that's not currently possible with k8s.
EDIT: forgot to mention that all this discussion matters only inasmuch you have strict latency targets. If you're ok to occasionally have very slow requests and your customers are ok with that and your bosses don't freak out because they spoke to that one customer that is very angry that their requests time out a few times every month ... I'm very happy for you! Not everybody has these requirements and you can go pretty far with a production system consisting of a couple of VM on AWS without k8s, containers, cgroups or whatever in the picture. I know people who do and it works just fine for them. But in order to understand what the fuss about all of this we need to frame this discussion as pertaining to batch vs controlled-latency. Otherwise it's hard to explain why there are so many otherwise intelligent people making choices that appear to be a bit silly, isn't it? Sure, there are people who are "just wrong" on the internet :-) but more often than not if somebody ends up with a different solution is because they have different goals and tradeoffs.
The best way to utilize resources is to overcommit, and the smart way to overcommit is to, say, allow 4x overcommit with each allocation limited to 1/4th of the available resources so not individual peak can choke the system. Given varied allocations, things average out with a reasonable amount of performance variability.
Idle CPUs are wasted CPUs and money out the window.
This assumes that the scheduled workloads are created equal which isn't the case. The app owners do not have control over what else gets scheduled on the node which introduces uncontrollable variability in the performance of what should be identical replicas and environments. What helps here is .. limits. The requests-to-limits ratio allows application owners to reason about the variability risk they are willing to take in relation to the needs of the application (e.g. imagine a latency-sensitive workload on a critical path vs a BAU service vs a background job which just cares about throughput -- for each of these classes, the ratio would probably be very different). This way, you can still overcommit but not by a rule-of-thumb that is created centrally by the cluster ops team (e.g. aim for 1/4) but it's distributed across each workload owner (ie application ops) where this can be done a lot more accurately and with better results. This is what the parent post is also talking about.
> This assumes that the scheduled workloads are created equal which isn't the case.
This particular allocation technique benefits from scheduled workloads not being equal as equality would increase likelihood of peak demand collisions.
When you set CPU reservations properly in your container definitions, they don't "crawl[] to a halt" when other containers demand more CPU than they did before. Every container is guaranteed to get the CPU it requested the moment it needs it, regardless of what other containers are using right up till that moment. So the key is to set the requested CPU to the minimum the application owner needs in the contended case. That's why I say that "requests are limits in reverse" - a request declared by container A effectively creates an implicit limit on container B when container A needs the CPU.
Perhaps you're concerned that not setting limits causes oversubscription on a node. But this is not so: when scheduling pods, the K8S scheduler only considers requests, not limits. If every container is forced to declare a CPU request - as it should - then K8S will not overprovision pods onto a node, and every container will be entitled to all the CPU it requested when needed.
Consider, too, that most applications are memory-bound and concurrency-bound, not CPU bound. And a typical web server is also demand-driven and load-balanced; it doesn't spontaneously consume CPU. For those kinds of applications, resource consumption is roughly consistent across all the containers, and the "pod running just find a moment ago...crawling to a halt" isn't a phenomenon you're going to see, as long as they haven't yet consumed all the CPU they requested. Limits hurt more than help especially for these kinds of workloads, because being able to burst gives them headroom to serve while a horizontal scale-out process (adding more pods) is taking place.
Most CPU-bound workloads tend to be batch workloads (e.g. ETL, map/reduce). Teams who run those on shared nodes usually just want to get whatever CPU they can. These jobs tend not to have strict deadlines and the owners don't usually panic when a worker pod has its CPU throttled on a shared node. And those who do care about getting all the CPU available are frequently putting these workloads on dedicated nodes, often even dedicated clusters. And the nodes tend to be ephemeral - launched just-in-time to run the job and then terminated when the job has completed.
As others have pointed out, putting CPU limits on containers is a contributing cause of nodes having lower CPU utilization than they could otherwise sustain. Every node costs money -- whether you bought it or rent it -- so underutilized nodes represent wasted capital or cloud spend. You want nodes to run as hot as possible to get the most out of your investment. If your CFO/cloud financial team isn't breathing down your neck when they see idle nodes, they ought to be.
Workloads that use up all the bursting with little benefit, or prioritizing burst volume for an HTTP server instead of a cronjob that'll finish in designated time anyway. We even had a case where developers weren't updating their requests while their app grew in requirements, and then had an incident on their hand when spare CPU time suddenly was sparse.
imagine you got 4 core ec2 instance, the latencies you see at 50% cpu utilisation and 90% cpu utilisation are quite different. with reservations similar thing happens i.e even though each container is guaranteed to their reservations, the relative CPU utilisation still increased very high when there are other busy processes on the same host
Although I saw this dissonance between actual and assigned CPU causing issues, particularly CPU throttling, I struggled to find a scalable solution that would affect all Go deployments on the cluster.
Getting all devs to include that autoprocs dependency was not exactly an option for hundreds of projects. Alternatively, setting all CPU request/limit to a whole number and then assigning that to a GOMAXPROCS environment variable in a k8s manifest was also clunky and infeasible.
I ended up just using this GOMAXPROCS variable for some of our more highly multithreaded applications which yielded some improvements but I’ve yet to find a solution that is applicable to all deployments in a microservices architecture with a high variability of CPU requirements for each project.
And as a maintainer of ko[1], it was a pleasant surprised to see ko mentioned briefly, so that's for that too :)
- "The JVM has been modified to be aware that it is running in a Docker container and will extract container specific configuration information instead of querying the operating system. The information being extracted is the number of CPUs and total memory that have been allocated to the container." https://www.oracle.com/java/technologies/javase/8u191-relnot...
- Here's a more detailed explanation and even a shared library that can be used to patch container unaware versions of Java. I wonder if the same could be done for Go?
"LD_PRELOAD=/path/to/libproccount.so java <args>"
https://stackoverflow.com/a/64271429
https://gist.github.com/apangin/78d7e6f7402b1a5da0fa3abd9381...
-
There are more recent changes to Java container awareness as well:
https://developers.redhat.com/articles/2022/04/19/java-17-wh...
Edit: or set -XX:-PreferContainerQuotaForCPUCount
I've been looking at going in the other direction and using a Go binary with a unikernel; the machine just runs the binary and nothing else. I haven't got this working to my satisfaction yet - it works but there's still too much infrastructure, and deployment is still "fun". But I think this is way more interesting for Go deployments than using containers.
Having said that, for serious business, this setup doesn't make sense. It possibly takes more work to operate as container when the gateway runs on a dedicated instance.
Of course there's also the well-known correlation between the quality of your k8s deployment and the number of proxies it hosts. /s
- Isolation of networking
- Isolation of process namespace
- Isolation of filesystem namespace
- CPU and Memory limit enforcement
- A near-universal format for packaging, distributing, and running applications, with metadata, and support for multiple architectures
- A simple method for declaring a multi-step build process
- Other stuff I don't remember
Go doesn't need a format for packaging - it's one file. It's becoming common practice to embed everything else into the binary. (side note: I haven't done this with env files yet, and tend to deploy them separately, but I don't see any reason why we don't do this and produce binaries targetted at the specific deployment environments. I might give it a go).
I kinda prefer makefiles for the build stuff, or even just a script. The whole process of creating a Docker instance, pushing source files to it, trigging go build and then pulling back the binary seems redundant; there's no advantage of doing this in a container over doing it on the local machine. And it's a lot faster on the local machine.
Talking to people, it appears to be as mholt said: everyone just does everything in containers so apparently we do this too.
> OSes are still better than container runtimes at managing competing processes on the same host.
OSes and container runtimes are the same thing
For a subset of OSes
And yes, setting up a separate VM for each instance of a process is perfectly feasible. That's what all this cloud business was about in the first place.