Am 23.10.2013 um 17:18 schrieb Les Mikesell <lesmikesell at gmail.com>: > On Wed, Oct 23, 2013 at 4:01 AM, Leon Fauster > <leonfauster at googlemail.com> wrote: >> Am 23.10.2013 um 07:52 schrieb James A. Peltier <jpeltier at sfu.ca>: >>> | i have a new setup where the htdocs directory for the webserver >>> | is located on a nfs share. Client has cachefilesd configured. >>> | Compared to the old setup (htdocs directory is on the local disk) >>> | the performance is not so gratifying. The disk is "faster" compared >>> | to the ethernet link but the cache should at least compensate this >>> | a bit. Do they exist more pitfalls for such configurations? >>> | >>> >>> The best thing to do with respect to NFS shares is to make extensive use of caching >>> in front of the web servers. This will hide the latencies that the NFS protocol will >>> bring. You can try to scale NFS through use of channel bonding or pNFS/Gluster but >>> setting up a reverse proxy or memcached instance is going to be your best bet to making >>> the system perform well. >> >> >> All web-frontends (multiple) have the filesystem caching already in >> place (bottom layer). The application uses a key-value-store in memory (top layer) to >> accelerate the webapp (php). Nevertheless the performance is not satisfying. I was looking >> at some caching by the httpd daemon (middle layer). Any experiences with such apache >> cache out there? > > What kind of throughput and latency are you talking about here? NFS > shouldn't add that much overhead to reads compared to disk head > latency and if you enable client caching might be considerably > faster. If you are writing over NFS you don't get the same options, > though and sync mounts are going to be slow. bonded (just for failover) interface with speed: 1000Mb/s duplex: Full ping gives me a round-trip echo packet with ~ 0.139 ms writes ~ 57,9 MB/s (dd test) reads ~ 59,7 MB/s (uncached), 3,9 GB/s (cached) (dd test) nfsstat -m output: rw,nosuid,noexec,relatime,vers=3,rsize=32768,wsize=32768,namlen=255,soft, \ nosharecache,proto=tcp,timeo=20,retrans=4,sec=sys,mountaddr=xxx.xxx.xxx.xxx, \ mountvers=3,mountport=102734,mountproto=tcp,fsc,local_lock=none,addr=xxx.xxx.xxx.xxx async is implicit. UDP would perform better but we voted for reliability but even TCP should perform a bit better, i guess :) the webapp caches also stat calls. so far -- LF