[snip]
worth going to the trouble of implementing nonblocking io into the file reads of our httpd (whose docuement tree is all nfs-mounted)... Given the inherent latencies in NFS vs. local disk, I'm hoping that I could only issue reads to filehandles that already have read-ahead data.
[snip]
Interesting question, and one worth investigating. I'll just offer that it was actually filers that helped dispel the notion that there are "inherent latencies in NFS vs. local disk". Watch what iostat shows you during periods of heavy I/O and you might be surprised - NetApps often have _lower_ latency than local disk, at least if iostat's "svc_t" is to be believed. :-) I've been curious to know if the I/O path through the NFS/network code is actually more highly tuned than UFS/SCSI (or FCAL) drivers, or if it really just all comes out a wash...
Years ago on an old Sun 4/670MP with a pair of VME SunNet Coprocessors (an old NFS accelerator board) and a 1MB Sbus Prestoserve we ran some benchmarks and tests to prove to a group of software engineers that having local disks in their machines was _not_ necessarily faster than running diskless; we had 30 ELC's on two shared 10base2 subnets running various diskless, dataless, and diskfull configurations and ran various tests with the development tools they used - heavy compilation, make, rcs, etc. In most cases, the fully-diskless results were *faster* than running with local disk, even when machines were configured to swap over the net rather than to local disk.
Of course, things have changed tremendously since 1992 :-) but the experience showed me that the network can often be as fast or faster than local resources... so when filers appeared on the scene a few years after that I was pretty much already in the NAS camp, even before the acronym had been coined. :-)
Speaking of which, back to my Oracle tests...
-- Chris
-- Chris Lamb, Unix Guy MeasureCast, Inc. 503-241-1469 x247 skeezics@measurecast.com