[SATLUG] AoE results

[Travis H.]

On Tue, Dec 05, 2006 at 04:02:02PM -0600, K. Spoon wrote:
> The big problem with ATA is that it foists management of the bus onto
> your CPU.  With SCSI, you generally have a dedicated processor on the
> card that can play traffic cop and all the CPU has to do is handle
> interrupts.

Those cards, like hardware RAID, cost a lot to design, and sooner or
later they become deccelerators rather than accelerators, because you're
Quad Xeon P4 has more than enough cycles to handle interrupts compared
to the 8MHz 80186 on the card (stats totally made up).

> A good demonstration of this is what happens when you are doing a large
> write to a filesystem on an IDE drive that has low memory.  After a
> certain point (about 256MB if you've got a drive with an 8MB cache) the
> kernel will attempt to push more data to the drive than it can handle.
> The kernel will call pdflush() and attempt to buffer stuff to system
> memory...

Interesting.

> but when that runs out, it'll try and use swap.

Why doesn't it just pause?

> If the swap
> is on the same drive as the one you're hammering, the system load will
> start climbing geometrically because the proc is trying to handle not
> only the I/O duty for the write, but for the swap activity as well.

Ugh, yeah, that's why I see systems with one IDE drive die with load
averages of 10 and others with the same specs can actually be responsive
with load averages of 140... I/O thrashing will kill a server much
faster than CPU monopolization.

> For the SATA drive on the workstation, yes.  For the PATA drive in the
> shelf server, no... all I did was hardset the BIOS to run the disks at
> PIO4.

Well, there's the main cause of the CPU load.

Of course it will take lots of CPU time if you send out a byte at a
time (okay, N bytes... I don't know how many PIO4 does)... that's one
interrupt per N bytes instead of one per DMA transfer... DMA can
handle moving sequential bytes over the bus just fine without your
CPU micromanaging :-)
-- 
"Cryptography is nothing more than a mathematical framework for
discussing various paranoid delusions." -- Don Alvarez
<URL:http://www.subspacefield.org/~travis/> -><-