Curious about performance

[gokr]

Hi folks!

I have a box at VM5, and performance felt outstanding - in fact so good I was wondering a bit. nbench-byte (ftp://ftp.tux.org/pub/tux/mayer/nbench-byte-2.2.2.tar.gz) reported on the 28th of june:

MEMORY INDEX : 9.685
INTEGER INDEX : 7.580
FLOATING-POINT INDEX: 15.295
Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38

Which is pretty darn fast - it is a tad faster than my Pentium-M 1.5Ghz laptop. Now I retried it yesterday (after the power failure incident) and I got Mem1.8, Int1.6, Flo3.4 which is... about 5 times slower! Ehm, and today... we are back to faster values again.

So my question is, what is the guaranteed performance I get? In other words - how much is a CPU "unit" or yet in other words - how many units do you put on each server (normally)?

[matta]

Hi,

To answer the last question first, you get 1 CPU unit per MB of memory. On a 4GB host this equates to 4096 total CPU units.

How long after the reboot did you run the benchmark? All the reboots were hard reboots as power was cut to the servers. Due to this all the RAID arrays have to do a full init upon boot which uses a lot of disk I/O for about 1-3 hours depending on the host load. I did set the rebuild rate to be for fastest I/O (as opposed to fastest rebuild), but the disks are still very busy.

At the same time as the init is being performed on boot, booting all the VM's themselves uses a lot of resources. In event of a cold reboot it does take at least 1 hour for everything to settle down to normal. On a soft reboot this is still about 30 minutes.

[gokr]

Hi!

Ok, so if each box is equipped with 4Gb RAM and 2 AMD CPUs (IIRC) then a 96-account would guarantee me 96/4000 = 2.4% of those two CPUs, correct? Worst case that is. Btw, perhaps this should be included in the FAQ - the number of units per server that is (or perhaps it already is, but I didn't see it).

Anyway, nbench today shows:

MEMORY INDEX : 9.130
INTEGER INDEX : 7.209
FLOATING-POINT INDEX: 14.357

Which is very good, I ran it at: Fri Jul 1 11:07:20 CEST 2005

That can't possibly be anywhere near the worst case, so I wonder a bit what the common "load" is. If the numbers above are the "normal" numbers then I am more than pleased!

regards, Göran

[matta]

W're up to vm9 and only once had a complaint about performance (see: vm6 feels sluggish). For the most part what you get now should get. Maybe someone with an account on vm1-vm4 will pot their nbench score as those servers have had quite a bit of time to generate load.

What is different with the unixshell# service than other "VPS" providers is:

1) We use Xen (Which rules!)
2) Many of our customers are the technical/geeky type. In general VPS world you sell to resellers who then cram as many websites as they can on a small VPS. Many of our customers use the account to test out distros, monitoring, DNS servers, development, etc which hardly is noticable to the overall system.

[revve]

For reference: vm4

MEMORY INDEX : 9.228
INTEGER INDEX : 7.236
FLOATING-POINT INDEX: 14.419