Recently we have the case of very high latencies on NFS reads as
reported by application (SAP R/3). NFS server was NetApp FAS; according
to NetApp statistic, average volume read latencies were in order 10ms,
while SAP stats gave 30-50ms. Systems were interconnected by dedicated
1Gb/s Cisco switches (3750G) with ca. 30% max load on interfaces.
On advice of my colleague we changed sunrpc.tcp_slot_table_entries from
default 16 to 128 which seemed to make situation much better - without
changing load pattern of filer in any visible way.
Now, I can understand, why we observed much higher latency on system and
why changing (what effectively is) queue depth helped. But I am totally
frustrated that there does not appear to be *any* possibility to detect
this situation on Linux side and to get a real numbers of real NFS IO
latencies or number of requests waiting to be executed (and I do not
even dream about per-mount point stats).
I am grateful for any hints how can we monitor Linux NFS client and get
real-life numbers of what happens inside. Thank you!
-andrey
On Sat, 2009-07-11 at 21:03 +0400, Andrey Borzenkov wrote:
> Recently we have the case of very high latencies on NFS reads as
> reported by application (SAP R/3). NFS server was NetApp FAS; according
> to NetApp statistic, average volume read latencies were in order 10ms,
> while SAP stats gave 30-50ms. Systems were interconnected by dedicated
> 1Gb/s Cisco switches (3750G) with ca. 30% max load on interfaces.
>
> On advice of my colleague we changed sunrpc.tcp_slot_table_entries from
> default 16 to 128 which seemed to make situation much better - without
> changing load pattern of filer in any visible way.
>
> Now, I can understand, why we observed much higher latency on system and
> why changing (what effectively is) queue depth helped. But I am totally
> frustrated that there does not appear to be *any* possibility to detect
> this situation on Linux side and to get a real numbers of real NFS IO
> latencies or number of requests waiting to be executed (and I do not
> even dream about per-mount point stats).
>
> I am grateful for any hints how can we monitor Linux NFS client and get
> real-life numbers of what happens inside. Thank you!
See the nfs-iostat utility in the nfs-utils package:
http://git.linux-nfs.org/?p=steved/nfs-utils.git;a=blob;f=tools/nfs-iostat/nfs-iostat.py;h=9626d42609b9485c7fda0c9ef69d698f9fa929fd;hb=HEAD
Trond
On Sunday 12 of July 2009 02:44:27 Trond Myklebust wrote:
> On Sat, 2009-07-11 at 21:03 +0400, Andrey Borzenkov wrote:
> > Recently we have the case of very high latencies on NFS reads as
> > reported by application (SAP R/3). NFS server was NetApp FAS;
> > according to NetApp statistic, average volume read latencies were
> > in order 10ms, while SAP stats gave 30-50ms. Systems were
> > interconnected by dedicated 1Gb/s Cisco switches (3750G) with ca.
> > 30% max load on interfaces.
> >
> > On advice of my colleague we changed sunrpc.tcp_slot_table_entries
> > from default 16 to 128 which seemed to make situation much better -
> > without changing load pattern of filer in any visible way.
> >
> > Now, I can understand, why we observed much higher latency on
> > system and why changing (what effectively is) queue depth helped.
> > But I am totally frustrated that there does not appear to be *any*
> > possibility to detect this situation on Linux side and to get a
> > real numbers of real NFS IO latencies or number of requests waiting
> > to be executed (and I do not even dream about per-mount point
> > stats).
> >
> > I am grateful for any hints how can we monitor Linux NFS client and
> > get real-life numbers of what happens inside. Thank you!
>
> See the nfs-iostat utility in the nfs-utils package:
>
> http://git.linux-nfs.org/?p=steved/nfs-utils.git;a=blob;f=tools/nfs-i
>ostat/nfs-iostat.py;h=9626d42609b9485c7fda0c9ef69d698f9fa929fd;hb=HEAD
>
This looks exactly like what is required, except ...
cn1:~ # ./nfs-iostat.py 2
Traceback (most recent call last):
File "./nfs-iostat.py", line 559, in ?
iostat_command(prog)
File "./nfs-iostat.py", line 464, in iostat_command
mountstats = parse_stats_file('/proc/self/mountstats')
File "./nfs-iostat.py", line 434, in parse_stats_file
f = file(filename)
IOError: [Errno 2] No such file or directory: '/proc/self/mountstats'
Which minimum kernel version is required for it?
Thank you!
On Sun, 2009-07-12 at 08:38 +0400, Andrey Borzenkov wrote:
> On Sunday 12 of July 2009 02:44:27 Trond Myklebust wrote:
> > On Sat, 2009-07-11 at 21:03 +0400, Andrey Borzenkov wrote:
> > > Recently we have the case of very high latencies on NFS reads as
> > > reported by application (SAP R/3). NFS server was NetApp FAS;
> > > according to NetApp statistic, average volume read latencies were
> > > in order 10ms, while SAP stats gave 30-50ms. Systems were
> > > interconnected by dedicated 1Gb/s Cisco switches (3750G) with ca.
> > > 30% max load on interfaces.
> > >
> > > On advice of my colleague we changed sunrpc.tcp_slot_table_entries
> > > from default 16 to 128 which seemed to make situation much better -
> > > without changing load pattern of filer in any visible way.
> > >
> > > Now, I can understand, why we observed much higher latency on
> > > system and why changing (what effectively is) queue depth helped.
> > > But I am totally frustrated that there does not appear to be *any*
> > > possibility to detect this situation on Linux side and to get a
> > > real numbers of real NFS IO latencies or number of requests waiting
> > > to be executed (and I do not even dream about per-mount point
> > > stats).
> > >
> > > I am grateful for any hints how can we monitor Linux NFS client and
> > > get real-life numbers of what happens inside. Thank you!
> >
> > See the nfs-iostat utility in the nfs-utils package:
> >
> > http://git.linux-nfs.org/?p=steved/nfs-utils.git;a=blob;f=tools/nfs-i
> >ostat/nfs-iostat.py;h=9626d42609b9485c7fda0c9ef69d698f9fa929fd;hb=HEAD
> >
>
> This looks exactly like what is required, except ...
>
> cn1:~ # ./nfs-iostat.py 2
> Traceback (most recent call last):
> File "./nfs-iostat.py", line 559, in ?
> iostat_command(prog)
> File "./nfs-iostat.py", line 464, in iostat_command
> mountstats = parse_stats_file('/proc/self/mountstats')
> File "./nfs-iostat.py", line 434, in parse_stats_file
> f = file(filename)
> IOError: [Errno 2] No such file or directory: '/proc/self/mountstats'
>
> Which minimum kernel version is required for it?
>
> Thank you!
You need kernel 2.6.17 or newer.
Cheers
Trond
On Sat, 2009-07-11 at 18:44 -0400, Trond Myklebust wrote:
> On Sat, 2009-07-11 at 21:03 +0400, Andrey Borzenkov wrote:
> > I am grateful for any hints how can we monitor Linux NFS client and get
> > real-life numbers of what happens inside. Thank you!
>
> See the nfs-iostat utility in the nfs-utils package:
>
> http://git.linux-nfs.org/?p=steved/nfs-utils.git;a=blob;f=tools/nfs-iostat/nfs-iostat.py;h=9626d42609b9485c7fda0c9ef69d698f9fa929fd;hb=HEAD
Oh, and recent versions of 'sar' (from the 'sysstat' package) should
also have support for NFS monitoring.
Trond