Wednesday, March 17, 2010

Performance tuning a server in less than three minutes while being slashdotted

So you wrote a blog post about something that seemed fairly innocuous, but for whatever reason, it caught the attention of one of the major sites and now your server load is at 110 and climbing, the ssh command line session is taking thirty seconds to respond to anything at all, and given that your post is on the front page of slashdot at primetime, this doesn’t look like it’s a temporary blip. What do you do?

Okay, first things first. You don’t have time to do a proper fine-tuning session. You need a quick & dirty tuneup here. Proper fine tuning you leave till after the traffic spike, and you can then come back at it with a plan and decent tools like siege and so on – but this is the “fix it in three minutes” version. So if you see stuff here that looks crude to you, despair not, everything looks crude when you’re trying to do it in that kind of timeframe.

First, you have to be able to actually use the server command line. The load’s at 110 – that’s not coming down anytime soon and until it does you can’t do the rest of the work you need to since it takes so much time for even a shell to respond. The load is being caused by Apache and MySQL taking on more work than the server can handle, causing the server to swap excessively; and you’ve got to dump that work or shut off the incoming new work to recover. You can try sudo killall -9 apache2 if you can get an open ssh terminal in to do the job (and it’s the first thing you should try), but the odds are that that server has to be reset. Whether that means a phone call to the data centre, or a walk down the hall to the server room, or just clicking a button in a web interface, that’s the first thing to do. Don’t hold off, because unless everyone stops reading your page right now, that load’s not coming down.

Once the box has rebooted (and I mean immediately – sit there watching ping until it comes back), ssh in and shut down apache. MySQL is okay for now, but the work is coming from Apache, so that has to get shut down until everything’s ready to go again. You’re going to lose a few minutes of service, yes, but that’s recoverable from for a blog (and if this is for something more serious than a blog, you’re going to be in trouble for not properly spec’ing the hardware to begin with anyway, so limit the damage now and take your licks later).

Read more: Stochastic Geometry

Posted via email from jasper22's posterous