[horde] performance when some components are not on local network

[Aleksandar Milivojevic]

Hi,

I'm having a problem with performance when some of the components  
(either IMAP server, or SQL database, or both) are not all on the same  
local network as web server where Horde is installed, but rahter on  
the other end of T1 based WAN.

For example, if web, IMAP and SQL servers are all on the local  
network, things work fast, almost instantly.

However, if IMAP server is not on the same network as web server,  
things slow down drastically.  Simmilary for SQL server on different  
network.  Takes some 20 seconds just for a user to log in.  So I took  
a look what is going on the network, why it takes that much time to  
simply log the user in.

I did tcpdump of traffic from the moment user clicks on "login"  
button, till the first page is fully loaded.  The dump contains some  
870KB of traffic between Horde, SQL and IMAP servers!!!  Just to  
generate a single page.  Large portions seem to be either duplicate  
data, or data that gets discarded/overwritten several times in a row.

Some examples.  The horde_sessionhandler table gets constantly updated  
again and again and again.  Wouldn't it be enough if it was updated  
only once, after the page is completely loaded?  Strangely, even with  
this constant updates, if I switch from SQL sessions to standard PHP  
sessions, time it takes to log the user in doesn't reduce noticably!?   
It could be that latency plays bigger problem than bandwith itself  
(since even in case of local PHP sessions, there's still relatively  
large number of queries going to SQL server).

The user's prefs seems to be loaded from database multiple times for  
single page load too.  These are not generating a lot of traffic, but  
you do it over and over again on network that has some latency, and it  
is definetely noticable.

Looking at the IMAP traffic, there's lot of not needed work there too.  
  There's not much traffic going in terms of data transferred.  But if  
you put the latency of WAN into calculation, it considerably slows  
down things.

Again some examples.  I can see IMP issuing some LIST commands to get  
list of folders.  It does LIST "" INBOX.%, and than repeats for other  
namespaces.  Than it does a new LIST for each returned folder.   
Hmmm...  Why doesn't it simply LIST "" INBOX.* and get all it needs in  
one command?  Or even LIST "" * and get really everything in one go?   
It also does bunch of LSUB commands, that basically don't reveal any  
new information that wasn't already present in output of LIST (or at  
least I'm failing to see it).

Anyhow, just wanted to point to some places where there might be a  
place to optimize things a bit (we are not all running gigabit  
networks, and even with those, the latency can be a killer).  Not to  
mention generating almost a megabyte of network traffic just to  
present a user with one simple page...  Now imagine 100 users doing it  
in parallel...

-- 
See Ya' later, alligator!
http://www.8-P.ca/