This past couple of days, I worked with our web guy as I redeployed our web cluster from FreeBSD 6.x-STABLE to FreeBSD 7.x-STABLE.
We also moved from the rock solid and never failing (and now end of lifed) PHP4 to PHP5 along the way.
My subjective view of FreeBSD 7 vs 6 on our web cluster gives me the feeling that 7 does feel faster with the new ULE scheduler.
I say subjective because I didn’t take real benchmarks before or afterwards, only on how the system felt (responsiveness) and using very basic tools (top, systat -vmstat).
As this was a cluster, I was able to do 1/2 of the machines at a time, so I was able to watch, for sake of argument, 2 systems handle the full load while on FreeBSD 6.
Then it was time to take down those servers and bring up the FreeBSD 7 systems. Outage was less than a couple seconds while the load balancer recomputed algorithms. Only reason it was measurable was because I really did take down the 6 systems while bringing up the 7 systems (in the load balancers). I did it this way because of the PHP differences and did not want to pollute session caches if there was a difference (I think it would be better to invalidate completely than to half-invalidate).
Whammo! We were up.
I waited a few minutes. I had already stopped the basic stats collection on the 6 systems just before removing from the cluster and started up stats collections after load was moved over. This would give me some side by side numbers to look at.
First, CPU utilization was down a full 3%. Now before you say ‘who cares?’, the systems were running ~92% idle, they now run closer to 95% idle and once I clean up the changes to the SNMP (should have done this right away, bad bad me), I’ll be able to see quantifiable items.
SNMP is also showing different memory usage, where under 6 the system used memory was close to 600-650MB resident, under 7 I am seeing 250-300MB resident. I don’t know where the change comes from at this point and will need to wait a few more days to determine the outcome. Could be a difference in PHP version, though I haven’t seen anything about PHP5 using less memory than PHP4. Can’t be an Apache difference as we stayed in the same version tree.
HTTP requests, though, didn’t change. At least not worth looking at as so far, my SNMP data shows a .002-.008 reduction in HTTP response time. I don’t think anyone is going to notice.
So that’s all in this weeks spew about computers.