When to add more servers?

[peng1can]

I've currently got two zimbra servers... one MTA that's mostly idle all the time and one mailbox server with 2x Dual Core Xeons and 8GB of RAM. I've got 7500 users and about 500 each of IMAP and Web sessions at peak times. Everyone I've spoken to at zimbra has said that this setup should be adequate, and that if I do outgrow it, I should pull the LDAP server off to a 3rd box. I received a lot of help in tweaking the system under 4.5 (NE) from a zimbra tech and the load on it was reasonable.

The problem is that ever since upgrading to 5.0 (NE), this box has a load average of around 6 during the day, and at times it jumps up around 22. The Java process is clearly the resource hog, and I've done my best to tune it according to the wiki. The zimbra tech assigned to my case appears to live in India, so he keeps looking at my server at 4am when the machine is idle and declaring that I don't have a problem, so I'm apparently on my own.

So I basically have two questions: Does 5.0 eat so many more resources that it's normal to see an increase in load after the upgrade? And if it is time to scale out, would I really see that big of a gain by splitting LDAP off to another machine, given that it's only using about 1% of the CPU time? Our server has been between sluggish and useless for the 10 days since the upgrade, so I need to do *something*. Any hints would be appreciated.

[bonadio]

Hello peng1can

I think you should investigate if the problem is CPU usage or DISK usage,
usually if you have a process doing a lot of disk io your load will go high
but your cpu usage will be low, run vmstat during the problem and
look for the last 2 columns CPU Id, CPU wa.

The CPU id will show how much of cpu is idle the higher the value less
CPU you are using.
The CPU wa will show how long the CPU is waiting for the disk, here higher
values means that you may have a disk that is slow for your load.

I think it is important to identify exactly what is causing the load, because
if you just add another hardware it may not solve the problem.

[]s