
Kernel NFS Server Statistics
============================

This document describes the format and semantics of the statistics
which the kernel NFS server makes available to userspace.  These
statistics are available in several text form pseudo files, each of
which is described separately below.

In most cases you don't need to know these formats, as the nfsstat(8)
program from the nfs-utils distribution provides a helpful command-line
interface for extracting and printing them.

All the files described here are formatted as a sequence of text lines,
separated by newline '\n' characters.  Lines beginning with a hash
'#' character are comments intended for humans and should be ignored
by parsing routines.  All other lines contain a sequence of fields
separated by whitespace.

/proc/fs/nfsd/pool_stats
------------------------

This file is available in kernels from 2.6.30 onwards, if the
/proc/fs/nfsd filesystem is mounted (it almost always should be).

The first line is a comment which describes the fields present in
all the other lines.  The other lines present the following data as
a sequence of unsigned decimal numeric fields.  One line is shown
for each NFS thread pool.

All counters are 64 bits wide and wrap naturally.  There is no way
to zero these counters, instead applications should do their own
rate conversion.

pool
	The id number of the NFS thread pool to which this line applies.
	This number does not change.

	Thread pool ids are a contiguous set of small integers starting
	at zero.  The maximum value depends on the thread pool mode, but
	currently cannot be larger than the number of CPUs in the system.
	Note that in the default case there will be a single thread pool
	which contains all the nfsd threads and all the CPUs in the system,
	and thus this file will have a single line with a pool id of "0".

packets-arrived
	Counts how many NFS packets have arrived.  More precisely, this
	is the number of times that the network stack has notified the
	sunrpc server layer that new data may be available on a transport
	(e.g. an NFS or UDP socket or an NFS/RDMA endpoint).

	Depending on the NFS workload patterns and various network stack
	effects (such as Large Receive Offload) which can combine packets
	on the wire, this may be either more or less than the number
	of NFS calls received (which statistic is available elsewhere).
	However this is a more accurate and less workload-dependent measure
	of how much CPU load is being placed on the sunrpc server layer
	due to NFS network traffic.

sockets-enqueued
	Counts how many times an NFS transport is enqueued to wait for
	an nfsd thread to service it, i.e. no nfsd thread was considered
	available.

	The circumstance this statistic tracks indicates that there was NFS
	network-facing work to be done but it couldn't be done immediately,
	thus introducing a small delay in servicing NFS calls.  The ideal
	rate of change for this counter is zero; significantly non-zero
	values may indicate a performance limitation.

	This can happen either because there are too few nfsd threads in the
	thread pool for the NFS workload (the workload is thread-limited),
	or because the NFS workload needs more CPU time than is available in
	the thread pool (the workload is CPU-limited).  In the former case,
	configuring more nfsd threads will probably improve the performance
	of the NFS workload.  In the latter case, the sunrpc server layer is
	already choosing not to wake idle nfsd threads because there are too
	many nfsd threads which want to run but cannot, so configuring more
	nfsd threads will make no difference whatsoever.  The overloads-avoided
	statistic (see below) can be used to distinguish these cases.

threads-woken
	Counts how many times an idle nfsd thread is woken to try to
	receive some data from an NFS transport.

	This statistic tracks the circumstance where incoming
	network-facing NFS work is being handled quickly, which is a good
	thing.  The ideal rate of change for this counter will be close
	to but less than the rate of change of the packets-arrived counter.

overloads-avoided
	Counts how many times the sunrpc server layer chose not to wake an
	nfsd thread, despite the presence of idle nfsd threads, because
	too many nfsd threads had been recently woken but could not get
	enough CPU time to actually run.

	This statistic counts a circumstance where the sunrpc layer
	heuristically avoids overloading the CPU scheduler with too many
	runnable nfsd threads.  The ideal rate of change for this counter
	is zero.  Significant non-zero values indicate that the workload
	is CPU limited.  Usually this is associated with heavy CPU usage
	on all the CPUs in the nfsd thread pool.

	If a sustained large overloads-avoided rate is detected on a pool,
	the top(1) utility should be used to check for the following
	pattern of CPU usage on all the CPUs associated with the given
	nfsd thread pool.

	 - %us ~= 0 (as you're *NOT* running applications on your NFS server)

	 - %wa ~= 0

	 - %id ~= 0

	 - %sy + %hi + %si ~= 100

	If this pattern is seen, configuring more nfsd threads will *not*
	improve the performance of the workload.  If this patten is not
	seen, then something more subtle is wrong.

threads-timedout
	Counts how many times an nfsd thread triggered an idle timeout,
	i.e. was not woken to handle any incoming network packets for
	some time.

	This statistic counts a circumstance where there are more nfsd
	threads configured than can be used by the NFS workload.  This is
	a clue that the number of nfsd threads can be reduced without
	affecting performance.  Unfortunately, it's only a clue and not
	a strong indication, for a couple of reasons:

	 - Currently the rate at which the counter is incremented is quite
	   slow; the idle timeout is 60 minutes.  Unless the NFS workload
	   remains constant for hours at a time, this counter is unlikely
	   to be providing information that is still useful.

	 - It is usually a wise policy to provide some slack,
	   i.e. configure a few more nfsds than are currently needed,
	   to allow for future spikes in load.


Note that incoming packets on NFS transports will be dealt with in
one of three ways.  An nfsd thread can be woken (threads-woken counts
this case), or the transport can be enqueued for later attention
(sockets-enqueued counts this case), or the packet can be temporarily
deferred because the transport is currently being used by an nfsd
thread.  This last case is not very interesting and is not explicitly
counted, but can be inferred from the other counters thus:

packets-deferred = packets-arrived - ( sockets-enqueued + threads-woken )


More
----
Descriptions of the other statistics file should go here.


Greg Banks <gnb@sgi.com>
26 Mar 2009