Re: [9fans] chess?

[erik quanstrom <quanstro@quanstro.net>]

> Thanks, this is possibly a way, but, at least in my case, this takes ages...
> (Say it can run for >45 mins; maybe much more; I do not know the true
> reason why,
> but I intuitively suspect the protocol.)
> Therefore I asked for a more specific pointer.
> Also, generally, there should be a quicker way.

it's easy to blame the protocol.  but it's more than that.
0.  the i/o unit is just 8k, this means that directories like
/sys/src/9 take 2 or more round trips.  this is trivial to fix
by upping MAXRPC in devmnt.  i've uppted this to
56k + IOHDRSZ.  this makes a difference even when the
rtt is 50µs.

1.  (or 0a) the kernel won't issue more than one concurrent
rpc for the same request.

2.  find/du are single threaded, so they they take full rtt latency
on each syscall.

there's a small demonstration of 2 at the end of this email.

- erik

---
note that except for the fcp test, rtt dominates.  even 384kbps
is more than enough to keep up.  in fact, except for the fcp
test time_link0 * rtt_link0/rtt_link1 tracks pretty closely with
the actual result.  since link1 is dsl, keeping track of the actual
rtt could have reduced some of this error.

link0:
rtt ~ 32ms uplink/downlink ~ 100mbps/100mbps:

0.29u 0.37s 895.45r 	 rc -c find | grep chess  # status= find 7086421: errors|
0.03u 0.03s 68.77r 	 rc -c cp lsr /tmp && grep chess /tmp/lsr
.0.04u 0.10s 5.40r 	 rc -c fcp lsr /tmp && grep chess /tmp/lsr

link1:
rtt ~ 52ms uplink/downlink ~ .384mbps/1.5mbps
0.00u 0.00s 1468.26r 	 rc -c find | grep chess
0.00u 0.01s 131.03r 	 rc -c cp lsr /tmp && grep chess /tmp/lsr >/dev/null
0.00u 0.00s 33.61r 	 rc -c fcp lsr /tmp && grep chess /tmp/lsr >/dev/null