How do we allocate the cpu we need?

[leftwo]

some questions im interested in

How do we allocate downstairs?

• one option: each downstairs gets up to X% cpu time allocated, independent of any other downstairs
  • two regions will not impact each others' available CPU time
  • may encounter throttling sooner than otherwise
• another option: multiple downstairs are grouped into one bucket of CPU time
  • If one VM is bursting IO, but others are not doing IO, that VM's downstairs gets full CPU time available to it
  • If multiple VMs burst together, they will slow each other down
    • may cause confusing behavior for operators investigating IO throttling

How do we allocate Upstairs

• Upstairs CPU is currently nontrivial for certain operations
  • Because upstairs is in the propolis process, we probably need to do any limiting at the tokio level
    • do we have any tools besides number of worker threads

related: #604

[faithanalog]

Per upstairs:

do we have any tools besides number of worker threads

I'm not sure we even have number of worker threads at the moment because most of my time observing upstairs with prstat has only one thread active at a time, and which thread is currently active just bounces around, so there's isn't really a whole lot of parallelism up there to clamp down on as far as I can tell.

Per downstairs:

The tradeoffs you mention are things ive been thinking about too. I think grouping all downstairs on a sled into one or more common CPU buckets on a sled seems like it would be ideal for general performance, but I think ideally an operator should be able to do one of

• allocate a VM they know will be doing a lot of IO to its own bucket
• prioritize a VM within a bucket
• at least choose which bucket the VM's disks go in

All of these require region allocation logic that we can't really do right now, so if we want to prioritize consistency then I think we should allocate a fixed amount of CPU per region.

also on the topic, worth mentioning that the region allocation logic right now, or at least the part of it that ive been touching while working with oxidecomputer/omicron#3416, does not limit the number of regions on a dataset. The only limits are whether the the dataset has room.

so if we allocate a fixed amount of CPU per region, we could still end up overloading a box by having way too many regions on it, but initially the work im doing to randomly distribute datasets across the rack should help make that problem less immediate

[askfongjojo]

Depends on having the cpu/memory limits framework