• G
    [PATCH] knfsd: knfsd: cache ipmap per TCP socket · 7b2b1fee
    Greg Banks 提交于
    Speed up high call-rate workloads by caching the struct ip_map for the peer on
    the connected struct svc_sock instead of looking it up in the ip_map cache
    hashtable on every call.  This helps workloads using AUTH_SYS authentication
    over TCP.
    
    Testing was on a 4 CPU 4 NIC Altix using 4 IRIX clients, each with 16
    synthetic client threads simulating an rsync (i.e.  recursive directory
    listing) workload reading from an i386 RH9 install image (161480 regular files
    in 10841 directories) on the server.  That tree is small enough to fill in the
    server's RAM so no disk traffic was involved.  This setup gives a sustained
    call rate in excess of 60000 calls/sec before being CPU-bound on the server.
    
    Profiling showed strcmp(), called from ip_map_match(), was taking 4.8% of each
    CPU, and ip_map_lookup() was taking 2.9%.  This patch drops both contribution
    into the profile noise.
    
    Note that the above result overstates this value of this patch for most
    workloads.  The synthetic clients are all using separate IP addresses, so
    there are 64 entries in the ip_map cache hash.  Because the kernel measured
    contained the bug fixed in commit
    
    commit 1f1e030b
    
    and was running on 64bit little-endian machine, probably all of those 64
    entries were on a single chain, thus increasing the cost of ip_map_lookup().
    
    With a modern kernel you would need more clients to see the same amount of
    performance improvement.  This patch has helped to scale knfsd to handle a
    deployment with 2000 NFS clients.
    Signed-off-by: NGreg Banks <gnb@melbourne.sgi.com>
    Signed-off-by: NNeil Brown <neilb@suse.de>
    Signed-off-by: NAndrew Morton <akpm@osdl.org>
    Signed-off-by: NLinus Torvalds <torvalds@osdl.org>
    7b2b1fee
svcsock.c 44.2 KB