Building Burstables: CPU slicing with cgroups
ubicloud.com
ubicloud.com
I've been a bit apoplectic in the past that cgroups seemed not super helpful in Kubernetes, but this really showed me how the different Kubernetes QoS levels are driven by similar juggling of different cgroups.
I'm not sure if this makes use of cpu.max.burst or not. There's a fun article that monkeys with these cgroups directly, which is neat to see. It also links to an ask that Kubernetes get support for the new (5.14) CFS Burst system. Which is a whole nother fun rabbit hole of fair share bursting to go down! https://medium.com/@christian.cadieux/kubernetes-throttling-... https://github.com/kubernetes/kubernetes/issues/104516
We chose not to use cpu.weight, and instead divide the host explicitly using cgroups (slice in systemd). We put Standard VMs in dedicated slices to keep them isolated and let several Burstable VMs share a slice. This provides a trade off between the price of the VM and resource guarantees.
We use cpu.max.burst to allow the VMs to "expand" a bit, while we understand that this creates a "noisy neighbor" problem. At the same time there is a minimum guarantee of the CPU. The cgroups allow for all those knobs and give a lot of control. Combining them in various ways is an interesting puzzle.
I run multi-tenant k8s clusters with hundreds of tenants and it fundamentally is a hard problem to balance workload performance with efficiency. Sharing resources increases efficiency but in most cases increases tail latencies.
1. Neighbours can be noisy to the other hyperthread on the same CPU. For example, heavy usage of avx-512 and other vectorized instructions can affect a tenant running on the same core but different hyperthread. You can disable hyperthreading, but now you are making the same tradeoff where you are sacrificing efficiency for low tail latencies.
2. There are certain locks in the kernel which can be exhausted by certain behaviour of a single tenant. For example, on kernel 5.15 there was one global kernel lock for cgroup resource accounting. If you have a tenant which is constantly hitting cgroup limits it increases lock contention in the kernel which slows down other tenants on the system which also use the same locks. This particular issue with cgroups accounting has been improved in later kernels.
3. If your latency sensitive service runs on the same cores which service IRQs, the tail latency can greatly increase when there are heavy IRQ load, for example high speed NIC IRQs. You can isolate those CPUs from the pool of CPUs offered to pods, but now you are dedicating 4-8 CPUs to just process interrupts. Ideally you could run the non-guaranteed pods on the CPUs which service IRQs, but that is not supported by kubernetes.
4. During full node memory pressure, the kernel does not respect memory.min and will reclaim pages of guaranteed QoS workloads.
5. The current implementation of memory QoS does not adjust memory.max of the burstable pod slice, so bursable pods can take up the entire free memory of the kubepods slice which starves new memory allocations from guaranteed pods.
Dont even get me started on NUMA issues.
This is fundamental - the closer utilization is to 100%, the higher the chance a newly arriving work item has to wait for one that's already running, and several already in the queue. What's astonishing is how steep that curve is. At 95% utilization the average queue length is 20 tasks. At 99% it's 100 tasks. At 99.9% it's 1000 asks. If you find yourself at 98% utilization, you should not think "nice - in fully utilizing the server I paid for" - you should buy another server and lower it to 49%. (Or optimize the code more)
One way to deal with this is to have separate low-latency and high-latency queues. You can then run low latency tasks at say 50% utilization and fill up idle time with high latency tasks. Presuming and you actually want the HL tasks to ever get done, you can't guarantee 100% utilization, but you can get arbitrarily close as long as there's high-latency work to do. I have no idea whether this is something Kubernetes can do. You can of course have more than two priority levels.
This applies everywhere there's a queue, which is basically everywhere there's s contended resource. Hyperscalers know this. It's even been theorized that S3 Glacier is just the super low priority disk access queue on regular AWS servers (but Amazon won't tell us).
So for instance if you run three processes for the same customer, can you set them to use the same cpu slices and deal with one of their apps occasionally needing a burst of CPU?
I came into the Linux world via Postgres, and this was an interesting project for me learning more about Linux internals. While cgroups v2 do offer basic support for CPU bursting, the bursts are short-lived, and credits don’t persist beyond sub-second intervals. If you’ve run into scenarios where more adaptive or sustained bursting would help, we’d love to hear about them. Knowing your use cases will help shape what we build next.
I particularly enjoyed the gentle exposition into the world of cgroups and how they work, the levers available, and finally how Ubicloud uses them.
Looking forward to reading how you handle burst credits over longer periods, once you implement that :)
Lovely work, Maciek!
The main challenge here is that cpu.max.burst can be set no higher than the limit set in cpu.max. This limits our options to some extent. But we can still look at some possible implementation choices here: - Pack more VMs into the same slice/group, and with that lower the minimum CPU guaranteed, and at the same time lower the price point. This would increase the chance of running into a "noisy neighbor", but we expect it would not be used for any critical workload. - Implement calculation of CPU credits outside of the kernel and change the CPU max and burst limits dynamically over an extended period of time (hours and days, instead of sub-second).
I imagine writing a custom scheduler would be quite an undertaking!
Are there typical use cases where you reach for cgroups directly instead of using the container abstraction?