Tweak nfs mount options for performance

[djeebus]

This should reduce network round trips, increase thread count

removed	added	effect
noac	actimeo=600	increase the attribute cache to 1 minute, fewer metadata operations
(default=cto)	nocto	disables close-to-open cache consistency, which slows down writes
rsize=524288	rsize=1048576	doubling the receive size means fewer tcp round trips, currently negotiated back down to 512 KB
(default=1)	nconnect=7	more connections increases parallel requests
lookupcache=none	lookupcache=pos	cache successful lookups of directory entries
(default=resvport)	noresvport	do not use a privileged port when making connections. increases maximum nfs mount points. unlikely to help, but recommended by google

---

Note

Optimizes nfs_mount_opts in iac/provider-gcp/nomad-cluster/main.tf to improve NFS throughput and reduce latency.

• IaC (GCP Nomad cluster):
  • Update locals.nfs_mount_opts in iac/provider-gcp/nomad-cluster/main.tf to optimize NFS behavior:
    • Add: actimeo=600, async, hard, lookupcache=positive, nconnect=7, nocto, noresvport, retrans=2, rsize=1048576, wsize=1048576, timeo=600, sec=sys.
    • Remove/replace: lookupcache=none, noac, tcp.
    • Keep: dynamic nfsvers, noacl, nolock.

^{Written by Cursor Bugbot for commit 2942fde. This will update automatically on new commits. Configure here.}

[ValentaTomas]

Is tweaking this to be in sync with our cca 4MiB reads make sense?

[djeebus]

Turns out that either our OS or the remote server is forcing us back down to 512 KB, so any increase gets ignored.

[ValentaTomas]

Interesting, any idea why?

[ValentaTomas]

@tomassrnka Any quick ideas?

[djeebus]

The documents say that the servers define a maximum. Different linux distros and NFS implementations use different values, this just happens to be Google Filestore's max.

[ValentaTomas]

Can

nconnect=7

have any negatives if we have a lot of client nodes?

[ValentaTomas]

Approving, would just like to resolve why the read/write sizes are being modified to the lower value.

[djeebus]

Can

nconnect=7

have any negatives if we have a lot of client nodes?

Google gives us 8,000 connections per TB, which is far more than we're currently thinking of using. Given that, I suspect there's not much of a down side here. Worth keeping in mind if we deploy this and we start seeing latency go up.