Clarify hardware recs for throughput #5472
Conversation
|
This change is |
There was a problem hiding this comment.
Reviewable status:
complete! 1 of 0 LGTMs obtained (waiting on @bdarnell and @rkruze)
jseldess
force-pushed
the
prod-checklist-update
branch
from
4edce5d to
e546174
Compare
| - The ideal configuration is 4-16 vCPUs, 8-64 GB memory nodes (2-4 GB of memory per vCPU). | ||
| - To optimize for throughput, use larger nodes, up to 16 vCPUs and 64 GB of RAM. Based on internal testing results, 16 vCPUs is the sweet spot for OLTP workloads. | ||
|
|
||
| To increase throughput further, add more nodes to the cluster instead of increasing node size; higher vCPUs will have NUMA](https://en.wikipedia.org/wiki/Non-uniform_memory_access)(non-uniform memory access) implications. |
There was a problem hiding this comment.
I'd remove this note about NUMA. We haven't explored this but we have no reason to believe it's a major concern at this point. The immediate reason why going to 32 vCPUs and beyond loses efficiency is simple mutex contention.
There was a problem hiding this comment.
Also for the record, on AWS m5/c5 nodes, you can go up to 48 vCPUs in a single NUMA group, so NUMA isn't a concern on this platform until you get to 72 vCPUs.
jseldess
force-pushed
the
prod-checklist-update
branch
from
e546174 to
289fbb2
Compare
jseldess
force-pushed
the
prod-checklist-update
branch
from
289fbb2 to
25c1407
Compare
There was a problem hiding this comment.
Reviewed 10 of 10 files at r2.
Reviewable status:
v19.1/deploy-cockroachdb-on-aws.md, line 72 at r2 (raw file):
- Run at least 3 nodes to ensure survivability. - Use `m` (general purpose), `c` (compute-optimized), or `i` (storage-optimized) [instances](https://aws.amazon.com/ec2/instance-types/), with SSD-backed [EBS volumes](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSVolumeTypes.html) or [Instance Store volumes](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ssd-instance-store.html). For example, Cockroach Labs has used `c5d.8xlarge` (36 vCPUs and 72 GiB of RAM per instance, NVMe SSD) for internal testing.
Our most-tested configuration is still the 4xlarge, i think.
v19.1/deploy-cockroachdb-on-google-cloud-platform.md, line 61 at r2 (raw file):
- Run at least 3 nodes to [ensure survivability](recommended-production-settings.html#topology). - Use `n1-standard` or `n1-highcpu` [predefined VMs](https://cloud.google.com/compute/pricing#predefined_machine_types), or [custom VMs](https://cloud.google.com/compute/pricing#custommachinetypepricing), with [Local SSDs](https://cloud.google.com/compute/docs/disks/#localssds) or [SSD persistent disks](https://cloud.google.com/compute/docs/disks/#pdspecs). For example, Cockroach Labs has used `n1-standard-32` (32 vCPUs and 60 GB of RAM per VM, local SSD) for internal testing.
Ditto, for n1-standard-16.
Increase optimal vCPUs to 32. Update GCP and AWS recs as well. Holding off on Azure until the upcoming cloud report, at which point we should revisit our hardware recs more broadly. Fixes #4711.
jseldess
force-pushed
the
prod-checklist-update
branch
from
25c1407 to
595b04c
Compare
Increase optimal vCPUs to 32.
Update GCP and AWS recs as well. Holding off on Azure until
the upcoming cloud report, at which point we should revisit
our hardware recs more broadly.
Fixes #4711.