1============================
  2Kernel NFS Server Statistics
  3============================
  4
  5:Authors: Greg Banks <gnb@sgi.com> - 26 Mar 2009
  6
  7This document describes the format and semantics of the statistics
  8which the kernel NFS server makes available to userspace.  These
  9statistics are available in several text form pseudo files, each of
 10which is described separately below.
 11
 12In most cases you don't need to know these formats, as the nfsstat(8)
 13program from the nfs-utils distribution provides a helpful command-line
 14interface for extracting and printing them.
 15
 16All the files described here are formatted as a sequence of text lines,
 17separated by newline '\n' characters.  Lines beginning with a hash
 18'#' character are comments intended for humans and should be ignored
 19by parsing routines.  All other lines contain a sequence of fields
 20separated by whitespace.
 21
 22/proc/fs/nfsd/pool_stats
 23========================
 24
 25This file is available in kernels from 2.6.30 onwards, if the
 26/proc/fs/nfsd filesystem is mounted (it almost always should be).
 27
 28The first line is a comment which describes the fields present in
 29all the other lines.  The other lines present the following data as
 30a sequence of unsigned decimal numeric fields.  One line is shown
 31for each NFS thread pool.
 32
 33All counters are 64 bits wide and wrap naturally.  There is no way
 34to zero these counters, instead applications should do their own
 35rate conversion.
 36
 37pool
 38	The id number of the NFS thread pool to which this line applies.
 39	This number does not change.
 40
 41	Thread pool ids are a contiguous set of small integers starting
 42	at zero.  The maximum value depends on the thread pool mode, but
 43	currently cannot be larger than the number of CPUs in the system.
 44	Note that in the default case there will be a single thread pool
 45	which contains all the nfsd threads and all the CPUs in the system,
 46	and thus this file will have a single line with a pool id of "0".
 47
 48packets-arrived
 49	Counts how many NFS packets have arrived.  More precisely, this
 50	is the number of times that the network stack has notified the
 51	sunrpc server layer that new data may be available on a transport
 52	(e.g. an NFS or UDP socket or an NFS/RDMA endpoint).
 53
 54	Depending on the NFS workload patterns and various network stack
 55	effects (such as Large Receive Offload) which can combine packets
 56	on the wire, this may be either more or less than the number
 57	of NFS calls received (which statistic is available elsewhere).
 58	However this is a more accurate and less workload-dependent measure
 59	of how much CPU load is being placed on the sunrpc server layer
 60	due to NFS network traffic.
 61
 62sockets-enqueued
 63	Counts how many times an NFS transport is enqueued to wait for
 64	an nfsd thread to service it, i.e. no nfsd thread was considered
 65	available.
 66
 67	The circumstance this statistic tracks indicates that there was NFS
 68	network-facing work to be done but it couldn't be done immediately,
 69	thus introducing a small delay in servicing NFS calls.  The ideal
 70	rate of change for this counter is zero; significantly non-zero
 71	values may indicate a performance limitation.
 72
 73	This can happen because there are too few nfsd threads in the thread
 74	pool for the NFS workload (the workload is thread-limited), in which
 75	case configuring more nfsd threads will probably improve the
 76	performance of the NFS workload.
 77
 78threads-woken
 79	Counts how many times an idle nfsd thread is woken to try to
 80	receive some data from an NFS transport.
 81
 82	This statistic tracks the circumstance where incoming
 83	network-facing NFS work is being handled quickly, which is a good
 84	thing.  The ideal rate of change for this counter will be close
 85	to but less than the rate of change of the packets-arrived counter.
 86
 87threads-timedout
 88	Counts how many times an nfsd thread triggered an idle timeout,
 89	i.e. was not woken to handle any incoming network packets for
 90	some time.
 91
 92	This statistic counts a circumstance where there are more nfsd
 93	threads configured than can be used by the NFS workload.  This is
 94	a clue that the number of nfsd threads can be reduced without
 95	affecting performance.  Unfortunately, it's only a clue and not
 96	a strong indication, for a couple of reasons:
 97
 98	 - Currently the rate at which the counter is incremented is quite
 99	   slow; the idle timeout is 60 minutes.  Unless the NFS workload
100	   remains constant for hours at a time, this counter is unlikely
101	   to be providing information that is still useful.
102
103	 - It is usually a wise policy to provide some slack,
104	   i.e. configure a few more nfsds than are currently needed,
105	   to allow for future spikes in load.
106
107
108Note that incoming packets on NFS transports will be dealt with in
109one of three ways.  An nfsd thread can be woken (threads-woken counts
110this case), or the transport can be enqueued for later attention
111(sockets-enqueued counts this case), or the packet can be temporarily
112deferred because the transport is currently being used by an nfsd
113thread.  This last case is not very interesting and is not explicitly
114counted, but can be inferred from the other counters thus::
115
116	packets-deferred = packets-arrived - ( sockets-enqueued + threads-woken )
117
118
119More
120====
121
122Descriptions of the other statistics file should go here.