On Sat, Jan 29, 2022 at 01:22:43PM -0700, Ted Bullock wrote:
> On 2022-01-29 6:17 a.m., Stefan Sperling wrote:
> > cvs does not use unveil, which makes it difficult to infer a common baseline.
> > Could you try your 'got status' NFS test with unveil disabled?
> 
> Alright, with unveil disabled, on nfs with an older sparc64 cpu.
> Slightly more than twice as performant. I'm quite surprised honestly.
> 
> $ time got status
>     4m26.37s real     0m21.44s user     0m50.38s system

Unveil performs internal directory walks which could translate to more
NFS traffic going over the wire. Even small request/response latencies
could add up to something significant with such a large tree.

There should still be ways to improve 'got status' and/or unveil
performance, of course. Trivial optimizations are unlikely to be
found at this point, but are not entirely out of the question.
More likely, going significantly faster would require some clever
tricks to somehow reduce the number of stat(2) calls.

With unveil disabled, does the NFS -l mount option now make a
significant difference?