[BlueOnyx:09304] Re: Heavy load during lots of httpd requests

[Chris Gebhardt - VIRTBIZ Internet]

Robert Fitzpatrick wrote:
> Thanks Chris. When you say more power, you mean move them to a dedicated
> server off the VPS or extend more CPU/RAM to the VPS? It didn't seem the
> RAM was an issue as it does not pull anything from Swap while under
> heavy load. The CPU seems to go under heavy load due to the number of
> httpd processes running, which in turn seem to be waiting on a response
> from something else.

I think I would look at 3 main areas:
1. RAM  (which you don't feel is the issue)
2. Processor
3. Network capacity

Also, since you mention VPS, add a 4th area:
4. Drive


With regard to the processor, keep in mind that with VPS you're getting 
a portion of whatever is available on the underlying hardware.  It's 
common practice to "oversubscribe" a virtualized host.  After all, 
that's the whole point of virtualization.

If your customer has spurts of high activity, and you have traffic & 
activity on the host from other customers, you could be pushing the 
limits of what the machine can process.

Also keep in mind that in about 90% of cases (in my experience, anyhow) 
drive access has been the first bottleneck for performance.  That is, 
with various VPS's running, the individual VPS access to the relatively 
low-bandwidth drive controllers is the slowdown.  This has been the case 
with several "low end" (ie: using internal SATA or even SAS drives) 
installations I have seen - even when CPU consumption is relatively 
minimal.  All those disk read/writes can really bog things down.

For maximum performance, I always recommend dedicated hardware.

With the network capacity, I don't know what your architecture looks 
like.  Our network core is capable of 192Gbps, but the effective 
available throughput for any given customer in our datacenter will be 
significantly less than that, depending on several factors.  Most of our 
customers have 100Mbps drops.  That leaves an effective 80-90Mbps of L3 
(TCP) available throughput once you account for typical overhead.  Just 
last week we had a customer with about half a rack of equipment sitting 
behind a 100Mbps switch that connected to the uplink we provide.  They 
were having an issue with packet-loss when their traffic approached 
about 45Mbps.   We helped them trace the issue to their switch, which 
was a consumer-grade un-managed switch.   As it turns out, the switch 
fabric was not robust enough to carry that traffic combined with their 
LAN traffic between servers and started dropping packets.  We handed 
them an old Cisco 2924 we had laying around and the problem instantly 
disappeared.   The moral of the story: it never hurts to investigate 
your network.

Best of luck.  I hope you're able to get things ironed out.
-- 
Chris Gebhardt
VIRTBIZ Internet Services
Access, Web Hosting, Colocation, Dedicated
www.virtbiz.com | toll-free (866) 4 VIRTBIZ