[PLUG] SMP big box vs. many small boxen for database / web server performance.

Aaron Burt aaron at bavariati.org
Sun Nov 7 05:36:05 UTC 2004


On Fri, Nov 05, 2004 at 01:32:47PM -0800, Randall Lucas wrote:
> 1. In a nutshell, for an Apache / Perl / MySQL web application with a 
> mix of DB reads and writes (assume there's a 1:5:2 http:dbread:dbwrite 
> ratio, very roughly), which of the two configurations below would, in 
> your opinion, give the best performance / value ratio?
> 	A. one web server and one database server, each with 1 unit 
> 	processor and 1 unit RAM;
> 	OR
> 	B. a single machine with 2 x 1 unit processor and 2 unit RAM?

B, since your DB traffic is higher than your net traffic, not that
it'll matter with real world bandwidth.  Also, a single box is easier
to admin and offers fewer points of failure.  *nix handles multiple
things on a single box far more gracefully than the competition.

Honestly, I doubt you'll need the dual proc, even.  Modern CPUs are
really, really fast.  Get it working and tested on a generic
single-proc 3+ GHz box with a good mobo, half gig of RAM and an IDE
drive, and then upgrade from there.  First suggestion would be
mirrored hotswap SCSI disks.

If you simply *must* get a box to stick in the colo right away, get a
2U rackmount from Dell (PE2650's good) or Supermicro.

> 2. And corollary to 1.:
> 	- if A is your choice, are there numbers that make B
> 	compelling?  In other words, what if B contained quad
> 	processors, or 4 units RAM?

If the web frontend was simple and lightweight, and the DB was huge
and handling complex queries.  In that case, you're basically spending
the money on a heavier DB server instead of a separate web server.

> 	- if the scenario is changed so that you are comparing 2 x (1
> 	unit RAM, 1 unit CPU) + 1 x (2 unit RAM, 2 unit CPU) versus 1
> 	x (4 unit ram, 4 x 1 unit CPU), does your choice remain the same?

That's too complex.  Don't make us do your homework for you, man.

> Thoughts?

You're overthinking the problem.  Also, I think you may have some
order-of-magnitude misconceptions of the relative bandwidths of system
buses, LANs, typical Internet connections and DB traffic.

Computer hardware gets cheaper, faster and better really quickly.
Buying just what you need now is generally the best strategy.





More information about the PLUG mailing list