RE: Elevator vs first-come-first-served

Robert Minichino (rmini@joni.pasture.net)
Tue, 17 Nov 1998 03:48:44 -0500 (EST)


> "can actually decrease performance"
> !!
> can fail to help, sure, but other than the (wasted) cost of sorting,
> how would it decrease performance? In general, that is, assuming some
> probabilistic model of requests. I suppose you could construct a
> Byzantine case.

When disk transaction frequency is at a high level in busy servers, the wasted
cost of sorting does become an issue; also, it is not extremely rare for the
OS to sort the data into the worst case for the controller's ordering
algorithm, as the OS knows nothing (especially with SCSI-SCSI RAID
controllers) about the true disk geometry. Being able to disable the sorting
adds little to no complexity and and affords a small, but significant,
increase in speed. Having more control over the sorting is even better. In
benchmarks on Solaris with raw disk devices and a benchmarking program that
simulated different ordering algorithms, throughput was highest to the disks
(behind a cached controller with 64MB) when first-come first-served was used
to service the randomly generated requests. I can rewrite and re-run the
benchmark if anyone desires.

--
Robert Minichino
Chief Engineer
Denarius Enterprises, Inc.

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/