Hey, You can also play with "/proc/sys/vm/swapiness" to avoid / limit swapping... But as explained, it's a bad idea to let a lot balancer swapping. It's supposed to introduce a very very low delay and swapping would increase that delay. Just ensure you have enough memory to handle the load you need/want.
cheers On Fri, Mar 18, 2011 at 7:33 PM, Ben Timby <bti...@gmail.com> wrote: > On Fri, Mar 18, 2011 at 2:00 PM, Antony <ddj...@mail.ru> wrote: >> Hi guys! >> >> I'm new to HAProxy and currently I'm testing it. >> So I've read this on the main page of the web site: >> "The reliability can significantly decrease when the system is pushed to its >> limits. This is why finely tuning the sysctls is important. There is no >> general rule, every system and every application will be specific. However, >> it is important to ensure that the system will never run out of memory and >> that it will never swap. A correctly tuned system must be able to run for >> years at full load without slowing down nor crashing." >> And now have the question. >> >> How do you usually prevent system to swap? I use Linux but solutions for any >> other OSes are interesting for me too. >> >> I think it isn't just to "swapoff -a" and to del appropriate line in >> /etc/fstab. Because some people say that it isn't good choise.. > > Prevent swapping by ensuring your resource limits (max connections) > etc. keep the application from exceeding the amount of physical > memory. > > Or conversely by ensuring that your physical memory is sufficient to > handle the load you will be seeing. > > This is what is referred to in the documentation, you need to tune > your limits and available memory for the workload you are seeing. Of > course simple things like not running other memory hungry applications > on the same machine apply as well. This is an iterative process > whereby you observe the application, make adjustments and repeat. You > must generate test load within the range of normal operations for this > tweaking to be true-to-life. Of course once you go into production the > tweaking will continue, no simulation is a replacement for production > usage. > > The reason running without swap is bad is because if you hit the limit > of your physical memory, the OOM killer is invoked. Any process is > subject to termination by the OOM killer, so in most cases decreased > performance is more acceptable than loss of a critical system process. > >
