[Techtalk] Server Deployment

[Dave North]

James:
> Well, I'm not making slashdot2.org, so I don't think it will be an
> insane amount of traffic. :)

I somehow sensed that from your original post.

> Whats the best way to 'estimate' future usage for request/second and
> such?

I don't think it will much matter. The question you're asking can be
answered fairly simply at this stage, since we already know the parameter
is "bang for the buck" and you're probably going to start with two servers
no matter what, right?
	Here's some practical advice:
1. PC133 RAM is almost as fast as the DDR in the real world for these
applications (in fact, I have not been able to measure a build/server
difference in my testing) ... BUT ... the cost is almost equivalent too.
	What this means is, at this point I wouldn't hinge my choice of
motherboard around either memory type. 512 chips are easier to get in
standard RAM, and can be had a little cheaper, but it won't make much
difference.*
	Another point about RAM: it's arguable that 3/4gig is better than
1gig, since you hit the "kernel barrier" at (approx) 980MB and take a bit
of a performance hit (I could barely see it, but Alan Cox states it's
anywhere from "nasty" to "horrible" under heavy load).

2. IDE is cheaper and nearly as fast as SCSI. ATA66 is fine, but ATA100 is
just as cheap now -- might as well use those. ATA133 is too cutting edge
for too little return for "bang for the buck" folks.

3. RAID will not significantly improve disk throughput in any measureable
way if you're running ATA66 or better. But it's handy to have, so you
might plan around it anyway. Does take up a little processor, but running
software raid costs nothing other than a little study.

4. There is no measureable difference in performance between P3 and AMD
chips at this time, save that the P3 will run cooler (and not involve some
bugs people have hit with AMDs). A comparable P3 motherboard will normally
be cheaper, so that might be an issue worth inspecting.
	The P4 might be worth considering, but not if you:

5. Use dual processors. This is very stable on Linux, and scales
wonderfully, so for performance/dollar, a slightly older dual motherboard
of reasonable quality will definitely fly.
	This was the single most important performance crank I measured in
all my testing.

Dave

*controversial comment:
        Another point about RAM: it's arguable that 3/4gig is better than
1gig, since you hit the "kernel barrier" at (approx) 980MB and take a bit
of a performance hit (I could barely see it, but Alan Cox said it's
anywhere from "nasty" to "horrible" under heavy load).
	So if you're not going to pack in 1.5gig (or better) it may be
that less is more.
	I have also noticed that most folks building server/performance
machines automatically pack the ram slots with far more ram than they
actually use (as if a bigger bucket will somehow magically carry water
_faster_).
	The more slots you fill, the slower your ram will run (again,
Alan Cox claims filling all three is much slower than just two; I find
this hard to measure but correct). The real point is, minimize your slots.
Start with 512MB on one board and see if your ram packs up. If it doesn't,
stop for now. If it does, add more RAM. Sounds reasonable, no?
	Cheaper too.