| gabriel rosenkoetter on Sun, 12 Jan 2003 09:43:04 -0500 |
|
On Sun, Jan 12, 2003 at 09:00:22AM -0500, Kam Salisbury wrote: > I think the quality and capabilities of the motherboard are most important > of all. And I still say that a workstation-level motherboard cannot accept ECC RAM. (Yes, it costs more, and yes, you don't *have* to use it, but the pain that comes from NOT being able to use it and wanting or needing it is unfixable. The fact that you can't get a rack-mount server from Dell with ECC RAM without stepping beyond the PowerEdge 2xx0 series was enough to make my employer switch to IBM as a vendor. Of course, IBM only has a single power supply in their xSeries 3xx systems--which are a royal pain to get mounted in an industry-standard 19" rack--so I'm not too sure where we're going next.) > The ability of the motherboard to accept a certain type of RAM or > processor or even storage device is not all that relevant if you do not > intend to use that feature or capability. If you've got a system where you need none of the assurances provided by redundant power supplies, UPS, ECC RAM, and RAIDed disk, then it's not a server in my book. :^> > Keep in mind, those Compaq, Dell and etc. servers that you may buy fall into > this same bucket. Unfortunately often true. And those machines shouldn't really qualify as servers. :^> Dell's lower-end PowerEdge servers lack some features they really ought to have (ECC RAM), as do IBM's lower-end eServers (redundant power supplies). Compaq has always and still does produce truly server-quality machines in their ProLiant line (two separate Mylex SCSI RAID controllers for the in-system bays, redundant power supplies); their weakness is that they only update their model line once a quarter (if that), and when they do, they don't jump straight to the newest stuff out, which means that they *still* aren't shipping systems with Xeons in them. > Once you begin talking about 32 way IBM boxes on AS400 then we > enter the realm of mainframe style computing. Barry Roomberg has an appropriate quote here: "mainframe style computing == not mainframe computing". ;^> > Beyond quality and support garuntees, you 'can' make a standard desktop > computer into a sub-server class machine. It really depends on how much > money you are willing to invest. You *can*, but if you're starting from scratch, right now, and you want a reliable system, buy one with the right stuff now, rather than having to replace effectively everything but the case later and you'll save money in the long run. > ECC RAM is great if you actually put that in (it costs more) and if the case > has redundant power supplies and you power the system with a clean UPS > device. I'd say those three things are orthogonal, though related. ECC RAM helps, not just for a power failure, but also if a DIMM develops errors under a certain threshold (and the way DIMMs are mass-marketed these days, they're liable to eventually). Redundant power supplies don't need to be plugged into separate UPSes to be useful; they can just be plugged into separate circuits (so if an air conditioner/hair iron/whatever blows one, the system doesn't go down). UPS's benefits for both regulating power (as a good--not just any, but a *good*--power strip will too) and providing time for the system to shut down cleanly (don't try to run off of them unless your UPS is actually uninteruptible... you'll know if it is, because you'll have installed a diesel generator) in the event of a true power loss. > both IDE and SCSI that beat the pants off of many competitors similiar > models. Yes... IDE RAID controllers. Yes... 5 disks. You really should check > their site. For a server with 25 to 50 people asking of it all day it is > enough. The money saved going the IDE route versus SCSI can be applied to > larger disks and a really good tape drive (USB or firewire of course, we > want to future proof ourselves right?). I don't agree. SCSI has always been a better protocol for many-disked machines, and it still is. I'm sure I know what 3ware's doing (using a controller per disk in their array), and I'm sure it's a bad idea (more moving parts to fail, and an expensive fix when they do, since 3ware's controllers are almost definitely not individually replaceable). One point I'm wavering on is the IDE-to-SCSI disk adapters. I'd never install one of these at work, but it's mighty attractive at home from the cost point of view. But I've read some reviews of the couple available, which showed that drives attached this way were noticeably both slower and more bursty than if they were just attached directly to an IDE bus, which sort of defeats the purpose, when you think about it. (In this case, there's still an extra moving part to fail, but a failure of the IDE controller attached to each disk doing the conversion to SCSI signals on the bus can more reasonably be viewed as a disk failure.) Oh, and if you're "future-proofing" yourself, USB (and even firewire) isn't where you want to be, fibre channel is. (Though precisely which kind of fibre-channel you want is a bit unclear.) > Need auto fail-over for network cards? Power supplies? Yep. they too can be > added to a PCI slot just like a really kickin' stereo to mom's mini-van. Sure, until you run out of externally-accessible PCI slots in that bargain-basement case, and have to replace it too. Also, how do you propose to run multiple power supplies out of a PCI slot that can all power the motherboard and communicate with each other? > I have and still actually do perform requests for just what I am talking > about. A client with an old box, lets say an old PII-333 workstation. Hrm. I seem to recall that the original question here was about purchasing a new computer, not about having an existing one to upgrade. If you were buying new parts, would you still by workstation-class materials and then upgrade them? Don't you think that would be more expensive? > harddisks, max out the RAM at 384MB and install a supplemental harddrive Nowhere *near* enough, in my book. Servers need to be running with, at a bare minimum, 512 MBs of RAM these days, especially if they're serving users with X sessions. And I wouldn't put anything new into production at work with less than 2 GBs. > install Redhat in a RAID1 configuration with Samba and a few extras. Viola'! Software RAID? Are you joking? I think the upgrades you do work for a certain category of user, wholly separate from the category of user I serve (who expect to spin all the processors at peak for days operating on several terabytes of data across NFS), which may be the real source of our difference in opinion on this. Despite this, my advice for someone purchasing new hardware remains: figure out what you need out of the system now, seriously considering features like ECC RAM, redundant power supplies, (hardware!) RAID, and (LVD!) SCSI rather than IDE for internal disks, and purchase accordingly. Buying a new workstation and then adding all the parts you need to it will quite probably result in redundant purchases (not of the power supply, but of the motherboard kind) and cost you way more in the long run. -- gabriel rosenkoetter gr@eclipsed.net Attachment:
pgpM4MtyJ7MV6.pgp
|
|