Date: Wed, 16 Apr 2003 10:52:08 -0400 From: David Gilbert <dgilbert@velocet.ca> To: "Michael Conlen" <meconlen@obfuscated.net> Cc: 'Alain Fauconnet' <alain@ait.ac.th> Subject: RE: tweaking FreeBSD for Squid using Message-ID: <16029.28184.11825.471228@canoe.velocet.net> In-Reply-To: <000a01c30423$bd0f2a40$2b038c0a@corp.neutelligent.com> References: <20030416024844.GC7867@ait.ac.th> <000a01c30423$bd0f2a40$2b038c0a@corp.neutelligent.com>
next in thread | previous in thread | raw e-mail | index | archive | help
>>>>> "Michael" == Michael Conlen <meconlen@obfuscated.net> writes: Michael> A lot of what you face doing a Squid server is backplane and Michael> other bus issues, though it's dependant on what you call Michael> "high performance" Michael> A pair of Sun E220R's (2 SPARC II processors) for example Michael> handled 1 million requests a day on a pair of mirrored 72 GB Michael> drives each. (Granted they were very nice 72GB drives). The Michael> thing about the Sun boxes was that they could get information Michael> out of memory really really fast, and the NIC cards could Michael> work to their full potential. Every device that did IO was on Michael> it's own PCI bus. There are several orders of magnitude in difference between motherboards (even of the same chipset) for PCI performance. PCI seems to be a bus that can be implemented well ... or very, very poorly. If you're planning to serve up 100Mbit plus from a PC, test several good (ie: expensive) motherboards in a bakeoff. Motherboards change so often that I can't even give you recomendations ... you can't buy them anymore. Ironically, many of the best motherboards for performance have also been high DOA. The K7S5A, for instance, had a DOA rate of 50% for us (50% crashed on memory stress tests, etc), but the good ones culled from the litter are among the best boards we have in production. Michael> It used to be that IDE drives took more processing power from Michael> the host to perform it's operations, where as SCSI does Michael> not. If that's still true I'd use that as a reason to stay Michael> away from IDE. The real advantage of SCSI (for large request rates) is tagged command queueing. Many spindles + tagged queueing = fast. Michael> The other advantage of SCSI, if you need great disk IO, is Michael> that you can have a lot of spindles. On a large SCSI system Michael> in a Sun for example I can get a single drive array to look Michael> like one SCSI device (with 14 disks in it) and put a lot of Michael> arrays on a channel. If I buy small, fast SCSI disks I can Michael> take full advantage of the 160 MB/sec array, where as I've Michael> seen a big fast IDE disk push no more than 10 MB/sec. The Michael> arrays can do RAID before it gets to the controller card, so Michael> you don't need the RAID in the box at all. RAID isn't always a win with Squid. Michael> Speaking of which, does anyone know of SCSI disk arrays with Michael> hardware RAID that work with FreeBSD? Michael> I've moved out of the Sun world and in to the FreeBSD world Michael> professionally and have no idea what's out there for PC Michael> hardware. As I've said before, in the category of non-silly-expensive RAID, vinum is faster than any I've tested. that said, SCSI<-->SCSI raid systems should all work with FreeBSD. Look in the hardware release notes for PCI raid devices, but dis-recomend them. Dave. -- ============================================================================ |David Gilbert, Velocet Communications. | Two things can only be | |Mail: dgilbert@velocet.net | equal if and only if they | |http://daveg.ca | are precisely opposite. | =========================================================GLO================
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?16029.28184.11825.471228>