Henrik Nordstrom disse na ultima mensagem: > tis 2007-03-20 klockan 20:13 +0545 skrev Tek Bahadur Limbu: > >> I admit there is no such rule but I am using it as a base for >> measurement and comparison. Obviously, req/sec is an easier and a >> better unit than say open_files/sec. > > What you want to monitor is the number of seeks/s, and if not that the > amount of time something is waiting for i/o. good point > > the first isn't very easily collected in most os:es, but the latter is > usually available via sar, iostat etc. I believe neither one is an effective method for measuring cache performance. What the deal is with caching is less bandwidth consumption. Speeding up network access is not so much the point anymore since we all have large bandwidth everywhere. So then what does it matter getting 1000 requests satisfied if each of them is 1k? I do compare the incoming http traffic to the outgoing. Higher the difference better my cache performance right. > > the big difference between ufs and aufs (and also diskd) is that with > aufs Squid does not wait while there is disk i/o, continuing network > operations as the disk i/o takes place. > > With ufs each millisecond spent in iowait means network activity was > paused.. > that is certainly an interesting point. IO Bound I guess can be fight by faster and cpu independent disks and subsystem (scsi) and then using polling on for example em (intel pro) nics which seem to "produce" less interrupts. Also setting vfs.write_behind and vfs.vmiodirenable may give important improvement on some hardware together with vfs.read_max. I do not know why net.isr.direct is not on by default but at least on SMP it is what you want. This probably still does not work well on older versions than 6.2 All this does not cut ufs's bottleneck but helps a lot. So sure diskd is the preferred cache_dir on FreeBSD. But again, not on low traffic machines where I can not find any difference. IMO so long as your machine does not handle more than 2mb/s it does not matter what you do FreeBSD does it well either way - supposed you have good hardware. Michel ... **************************************************** Datacenter Matik http://datacenter.matik.com.br E-Mail e Data Hosting Service para Profissionais. ****************************************************
