diff options
author | Shehjar Tikoo <shehjart@gluster.com> | 2011-01-11 01:40:16 +0000 |
---|---|---|
committer | Anand V. Avati <avati@dev.gluster.com> | 2011-01-11 23:58:33 -0800 |
commit | 89cbbee063243d84dc64b15b2e2f403be1ca226b (patch) | |
tree | 9ade31b94ccea56ebc88adf5702541eacfcb4cfb | |
parent | e344d58762e12aa1edd1d4b133e1c1f77229b370 (diff) |
nfs: Introduce tunable for memory consumption
NFS has used a common define called the mem-factor to determine
how much memory is allocated for the following items:
o inode table
o local structures used by nfs for storing fop state
o local structures used by nfs3 for storing nfs op state
These are all allocated through the mem-pool. The factor is multiplied by
a multiple that is specific to each data structure. For eg.
define GF_NFS_CONCURRENT_OPS_MULT 15
define GF_NFS_INODE_LRU_MULT 6000
The first value is used for allocating a mem-pool for storing
state for each fop or nfs op that is currently being handled. Knowing that linux
allows at most 128 in-flight requests, this multiple combined with the default
mem-factor of 15 gives us 225 slots in the mem-pool for the local structures.
Similarly, 6000*15 gives us a space of 90k inodes in the lru.
That means, increasing the common mem-factor will allow increasing the perf
under some conditions. This patch introduces the mem-factor as a configurable
option.
Signed-off-by: Shehjar Tikoo <shehjart@gluster.com>
Signed-off-by: Anand V. Avati <avati@dev.gluster.com>
BUG: 2277 (Regression in Gluster NFS re-read performance)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=2277
-rw-r--r-- | xlators/nfs/server/src/nfs.c | 28 |
1 files changed, 27 insertions, 1 deletions
diff --git a/xlators/nfs/server/src/nfs.c b/xlators/nfs/server/src/nfs.c index 979fd2fb04d..f7def106005 100644 --- a/xlators/nfs/server/src/nfs.c +++ b/xlators/nfs/server/src/nfs.c @@ -493,6 +493,22 @@ nfs_init_state (xlator_t *this) } nfs->memfactor = GF_NFS_DEFAULT_MEMFACTOR; + if (dict_get (this->options, "nfs.mem-factor")) { + ret = dict_get_str (this->options, "nfs.mem-factor", + &optstr); + if (ret < 0) { + gf_log (GF_NFS, GF_LOG_ERROR, "Failed to parse dict"); + goto free_foppool; + } + + ret = gf_string2uint (optstr, &nfs->memfactor); + if (ret < 0) { + gf_log (GF_NFS, GF_LOG_ERROR, "Failed to parse uint " + "string"); + goto free_foppool; + } + } + fopspoolsize = nfs->memfactor * GF_NFS_CONCURRENT_OPS_MULT; /* FIXME: Really saddens me to see this as xlator wide. */ nfs->foppool = mem_pool_new (struct nfs_fop_local, fopspoolsize); @@ -897,6 +913,16 @@ struct volume_options options[] = { .description = "Use this option on systems that need Gluster NFS to " "be associated with a non-default port number." }, - { .key = {NULL} }, + { .key = {"nfs.mem-factor"}, + .type = GF_OPTION_TYPE_INT, + .description = "Use this option to make NFS be faster on systems by " + "using more memory. This option specifies a multiple " + "that determines the total amount of memory used. " + "Default value is 15. Increase to use more memory in " + "order to improve performance for certain use cases." + "Please consult gluster-users list before using this " + "option." + }, + { .key = {NULL} }, }; |