Date: Tue, 11 Mar 97 10:01:33 CST From: Joe Greco <jgreco@solaria.sol.net> To: neal@pernet.net (Neal Rigney) Cc: mmead@goof.com, msmith@atrad.adelaide.edu.au, julian@whistle.com, tom@sdf.com, isp@freebsd.org, hackers@freebsd.org Subject: Re: freebsd as a news server? Message-ID: <199703111601.KAA27975@solaria.sol.net> In-Reply-To: <Pine.BSF.3.91.970308232838.13584B-100000@jennifer.pernet.net> from "Neal Rigney" at Mar 8, 97 11:35:14 pm
next in thread | previous in thread | raw e-mail | index | archive | help
> > Hmm, the disks are far from slow. They're 9G micropolis > > SCSI-II fast, "ultra wide" disks. There is, however, one bus. > > BIG problem there! If I remember Joe Greco's advice, you should have AT > LEAST two/three busses. I put history and spool on seperate busses(I'm > poor, I can't afford 15 SCSI controllers :) Actually, it's not a problem (from a certain point of view)... but it's a real waste. You don't NEED ultra wide on a news server, you're not transferring enough data to warrant it. My basic philosophy has always been to reduce resource contention whereever I can. The nice ultra wide SCSI controller that Matthew has probably cost $200-$300. I can outfit a system with three NCR-810 controllers for that price (actually more like $120 if I use the real cheap ones, a bit risky) and I gain three independent SCSI busses. I usually like to stripe _across_ the controllers, although I can't say for sure that putting spool and history on separate SCSI busses is a bad idea. The real idea is simply to provide as much independent I/O capacity as possible. > > There's only 5 drives on it. I'm interested in figuring out > > what's wrong with the layout. I've got the striping factor set > > to 255 blocks (per a suggestion in the docs for ccd). I'll try > > I've got the interleave set to 1k. I fiddled with the setting a little, > and found 1k appeared to give the best performance. If I remember > correctly, you want the "average" article to fit in one interleave. I > may be way off on that though. With the interleave set to 1k, we're able > to keep up with a releatively full feed (we don't get de.* and a couple > non-english hierarchies). We've got SCSI time to burn it appears. No > problem receiving the feed. Note, however, that we don't have many > active feeds out(ok, only 1 "full" feed), and we have a relatively low > reader count(we max out at about 45). On the other hand, we don't > anywhere NEAR enough drive space(5G now, adding 4 next week). But > overall, I don't have any problems. Our response time is good, and the > system's not hiccupped in about 6 months. BTW: We're running 2.1.6 > now. I don't think 2.1.6 does noatime though(maybe I'm just being dense > here?) Try vastly raising your stripe size, up to a cylinder group. You should see no decrease in performance, and perhaps an increase. This gets back into my whole lecture series about concurrency within an I/O system: basically you don't want an ARTICLE to fit in an interleave, you want the ARTICLE plus the DIRECTORY information to fit in an interleave. This way, when you go to read an article, you want to affect only one drive so that the other drives can be doing other things. ... JG
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199703111601.KAA27975>