[ Date Index ] [ Thread Index ] [ <= Previous by date / thread ] [ Next by date / thread => ]
On 24/03/14 23:25, Martijn Grooten wrote: > > I have to say, the performance improvements were a little disappointing. > Which probably shows I'm still doing something seriously wrong > somewhere. If there is a small working set and the writes are not blocking, then it is probably in RAM anyway, so tmpfs is going to have little benefit, and potentially negative since you probably made a suboptimal split compared to the kernel. I'd always benchmark such a system with bonnie++, and compare that to expected performance of the IO system, just to sanity check it, but beyond that you want specific ideas of what is slow. Ponder the rationale behind design of Varnish, your kernel does know how to efficiently read files from disk, and write them to disk, it might not be tuned to your specific case but it is probably good enough. It probably worth thinking about how/what you are doing, rather than how the system is configured. Unless you know already you are doing something silly. Only times I've seen IO bottleneck recently on even fairly low end modern hardware has nearly always been when it is rewriting files in a loop. e.g. Write out file with data A in, modify A, write out whole file again, rinse repeat. Sure there are situations where you know it will be a lot of data, but these nearly always from grubby low level systems, be it instrument data, or network monitoring, and often here it pays to ask - do I need everything. -- The Mailing List for the Devon & Cornwall LUG http://mailman.dclug.org.uk/listinfo/list FAQ: http://www.dcglug.org.uk/listfaq