On Fri 28 Jan, 2000, tkaczma@gryf.net wrote:

On Fri, 28 Jan 2000, Jason Middlebrooks wrote:

NetApp person. Besides, I though they were 36GB drives. As per performance, of course, there is bus arbitration overhead to start. Also

Bus arbitration may be reduced by reducing the number of active devices on a loop? At least that would be my intuition. Has anyone quantified any effects this might carry through to the head's performance? Analytic, simulated or empiracle evidence would all be acceptable!

you're talking about putting in huge drives, which maximum input/output is still about the same as the smaller drives. Drive performance hasn't improved all that much while capacity has increased significantly.

That's always been one of my concerns - in fact I think this has cropped up on the list before at least once.

Each successive volume increment in SCSI disks seemed to about double capacity while only raising IO's/s by say 20% over each generation. As capacities grow this gradient mismatch between the two characteristics means data gets relatively less accessible per spindle. This is why *more* spindles are often recommended for high-performance/bandwidth requirements. So far so orthodox.

It'd be nice to see what bus arbitration, RAID calculation overheads and fs-layout&maintenance take away by counterpoint due to such increased spindle-counts. If anything. My intuition is that spindles are a key factor in these algorithms, whereas larger linear arrays of sectors at each spindle are less significant. Any volunteers to argue either way?

Recently I've been pleasantly surprised by the performance of Seagate/Sun 9GB disks - able to transfer 2-3x what old 4.5GB disks could, but then I suppose this is a generation or two further on and 4.5GB disks aren't really being produced. Fibre-channel product almost certainly gets highest-performance controllers, interfaces and firmware so maybe Tom and I do them a disservice by suspecting them of the same capacity-curve mismatch described above.

Evidence about all this would be welcome. I'd bet we've almost all of us, at some point, bought at least one filer intended for heavy use (usenet, Oracle, etc.) and had to intuit how our disk-size, spindle-count and loop/chain-count decisions would impact our actual performance. Often this might even override questions of money (perish the thought!).

If someone in NetApp engineering or testing hasn't had their own curiosity piqued by this vexed question yet and drawn up matrices describing the tradeoffs, or even shown convincingly that there are no real tradeoffs to be made, then perhaps someone might sound them out about it soon?

Ever living in hope..

-- End of excerpt from tkaczma@gryf.net