Hi!
I need to tune a nfs server and client. on the server we have
several Tbyte of ~2Mbyte files and we need to transfer them read
only to the client. latency and throughput are crucial.
What nfs server should i use? i started with the
nfs-kernel-server on top of a kernel 2.6.22 on debian on the
server side. the client is a debian etch server (2.6.18 kernel)
with 1Gbyte e1000 intel network driver. later on we consider two
network cards on both machines to transfer 2Gbit/s. Jumboframes
are an option (how much will they help?)
Right now i have only four disks in the server and i get 50Mbyte
out of each of them, simultaniously, for real world loads (random
reads across the disk, trying to minimizing the seeks by reading
the files in one go with
for i in a b h i ; do ( find /var/disks/sd$i -type f | xargs -I=B0 dd if=3D=
=B0 bs=3D2M of=3D/dev/null status=3Dnoxfer 2>/dev/null & ) ; done
so with this (4*50 Mbyte/s) i should be able to saturate both
network cards.
accessing the disks with apache2-mpm-worker we get ~90Mbyte/s out
of the server, partly with considerable latency in the order of
magnitude of 10s.
I was hoping to get at least the same performance with much
better latency with nfs.
on the server i start 128 nfs servers (RPCNFSDCOUNT=3D128) and export
the disks like this:
/usr/sbin/exportfs -v
/var/disks/sda <world>(ro,async,wdelay,root_squash,no_subtree_check,anonuid=
=3D65534,anongid=3D65534)
/var/disks/sdb <world>(ro,async,wdelay,root_squash,no_subtree_check,anonuid=
=3D65534,anongid=3D65534)
/var/disks/sdh <world>(ro,async,wdelay,root_squash,no_subtree_check,anonuid=
=3D65534,anongid=3D65534)
/var/disks/sdi <world>(ro,async,wdelay,root_squash,no_subtree_check,anonuid=
=3D65534,anongid=3D65534)
on the client i mount them like this:
lotta:/var/disks/sda on /var/disks/sda type nfs (ro,hard,intr,proto=3Dtcp,r=
size=3D32k,addr=3D217.213.5.44)
lotta:/var/disks/sdb on /var/disks/sdb type nfs (ro,hard,intr,proto=3Dtcp,r=
size=3D32k,addr=3D217.213.5.44)
lotta:/var/disks/sdh on /var/disks/sdh type nfs (ro,hard,intr,proto=3Dtcp,r=
size=3D32k,addr=3D217.213.5.44)
lotta:/var/disks/sdi on /var/disks/sdi type nfs (ro,hard,intr,proto=3Dtcp,r=
size=3D32k,addr=3D217.213.5.44)
but when i then do the same dd again on the client i get
disappointing 60-70Mbyte/s altogether. from a single disk i get
~25Mbytes/s on the client side.
i played with some buffers /proc/sys/net/core/rmem_max and
/proc/sys/net/core/rmem_default and increased them to 256M on the
client.
i was suspecting that the nfs server reads the files in too small
chunks and tried to help it with =
for i in a h i ; do ( echo $((1024*6)) > /sys/block/sd$i/queue/read_ahead=
_kb ) ; done
to get it to read in the files in one go.
I would hope to at least double the speed. do
you have a benchmark tool that can tell me the latency? i tried
iozone and tried forcing it to only do read tests and did not get
any helpfull error or output at all. =
on the server:
nfsstat
Server rpc stats:
calls badcalls badauth badclnt xdrcall
98188885 0 0 0 0
Server nfs v3:
null getattr setattr lookup access readlink
5599 0% 318417 0% 160 0% 132643 0% 227130 0% 0 =
0%
read write create mkdir symlink mknod
97256921 99% 118313 0% 168 0% 0 0% 0 0% 0 =
0%
remove rmdir rename link readdir readdirplus
162 0% 0 0% 0 0% 0 0% 0 0% 105556 =
0%
fsstat fsinfo pathconf commit
0 0% 1270 0% 0 0% 7153 0%
cat /proc/net/rpc/nfsd
rc 0 118803 98069945
fh 0 0 0 0 0
io 3253902194 38428672
th 128 10156908 1462.848 365.212 302.100 252.204 311.632 187.508 142.708 14=
2.132 198.168 648.640
ra 256 97097262 0 0 0 0 0 0 0 0 0 64684
net 98188985 16 98188854 5619
rpc 98188885 0 0 0 0
proc2 18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
proc3 22 5599 318417 160 132643 227130 0 97256921 118313 168 0 0 0 162 0 0 =
0 0 105556 0 1270 0 7153
proc4 2 0 0
proc4ops 40 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0=
0 0 0 0 0 0 0 0
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
Hello Andreas,
On Thursday 25 October 2007 15:10:29 Andreas Schuldei wrote:
> lotta:/var/disks/sda on /var/disks/sda type nfs
> (ro,hard,intr,proto=tcp,rsize=32k,addr=217.213.5.44) lotta:/var/disks/sdb
try to increase rsize and wsize as much as possible, the maximum can be
adjusted in /proc/fs/nfsd/max_block_size on the nfs server.
>
> but when i then do the same dd again on the client i get
> disappointing 60-70Mbyte/s altogether. from a single disk i get
> ~25Mbytes/s on the client side.
>
> i played with some buffers /proc/sys/net/core/rmem_max and
> /proc/sys/net/core/rmem_default and increased them to 256M on the
> client.
>
> i was suspecting that the nfs server reads the files in too small
> chunks and tried to help it with
>
> for i in a h i ; do ( echo $((1024*6)) >
> /sys/block/sd$i/queue/read_ahead_kb ) ; done
Are your partitions on lvm or md? This wouldn't help then, AFAIK for lvm you
can't do it via /sys, but must always use blockdev
>
> to get it to read in the files in one go.
>
> I would hope to at least double the speed. do
> you have a benchmark tool that can tell me the latency? i tried
> iozone and tried forcing it to only do read tests and did not get
> any helpfull error or output at all.
First of all, what is you *local* performance?
Cheers,
Bernd
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
* Bernd Schubert ([email protected]) [071025 15:54]:
> Hello Andreas,
>
> On Thursday 25 October 2007 15:10:29 Andreas Schuldei wrote:
> > lotta:/var/disks/sda on /var/disks/sda type nfs
> > (ro,hard,intr,proto=tcp,rsize=32k,addr=217.213.5.44) lotta:/var/disks/sdb
>
> try to increase rsize and wsize as much as possible, the maximum can be
> adjusted in /proc/fs/nfsd/max_block_size on the nfs server.
the with a max_block_size and rsize=512k i did not achive a noticable performance impact.
increasing it to 2M decreased the performance to 12-12Mbytes/s per disk. do i need to adjust more
tuneables to benefit from the increased size?
> > i was suspecting that the nfs server reads the files in too small
> > chunks and tried to help it with
> >
> > for i in a h i ; do ( echo $((1024*6)) >
> > /sys/block/sd$i/queue/read_ahead_kb ) ; done
>
> Are your partitions on lvm or md? This wouldn't help then, AFAIK for lvm you
> can't do it via /sys, but must always use blockdev
they are on xfs directly on disk, no lvm or md involved.
> > to get it to read in the files in one go.
> >
> > I would hope to at least double the speed. do
> > you have a benchmark tool that can tell me the latency? i tried
> > iozone and tried forcing it to only do read tests and did not get
> > any helpfull error or output at all.
>
> First of all, what is you *local* performance?
i mentioned that i get 50Mbyte/s out of each disk, even
when doing the dd on all four disks at the same time.
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
On Oct 25, 2007, at 9:10 AM, Andreas Schuldei wrote:
> Hi!
>
> I need to tune a nfs server and client. on the server we have
> several Tbyte of ~2Mbyte files and we need to transfer them read
> only to the client. latency and throughput are crucial.
>
> What nfs server should i use? i started with the
> nfs-kernel-server on top of a kernel 2.6.22 on debian on the
> server side. the client is a debian etch server (2.6.18 kernel)
> with 1Gbyte e1000 intel network driver. later on we consider two
> network cards on both machines to transfer 2Gbit/s. Jumboframes
> are an option (how much will they help?)
>
> Right now i have only four disks in the server and i get 50Mbyte
> out of each of them, simultaniously, for real world loads (random
> reads across the disk, trying to minimizing the seeks by reading
> the files in one go with
>
> for i in a b h i ; do ( find /var/disks/sd$i -type f | xargs -I=B0 dd =
> if=3D=B0 bs=3D2M of=3D/dev/null status=3Dnoxfer 2>/dev/null & ) ; done
>
> so with this (4*50 Mbyte/s) i should be able to saturate both
> network cards.
>
> accessing the disks with apache2-mpm-worker we get ~90Mbyte/s out
> of the server, partly with considerable latency in the order of
> magnitude of 10s.
>
> I was hoping to get at least the same performance with much
> better latency with nfs.
With a single client, you should not expect to get any better =
performance than by running the web service on the NFS server. The =
advantage of using NFS under a web service is that you can =
transparently scale horizontally. When you add a second or third web =
server that serves the same file set, you will see an effective =
increase in the size of the data cache between your NFS server's =
disks and the web servers.
But don't expect to get better data throughput over NFS than you see =
on your local NFS server. If anything, the 10s latency you see when =
the web server is on the same system with the disks is indicative of =
local file system configuration issues.
> on the server i start 128 nfs servers (RPCNFSDCOUNT=3D128) and export
> the disks like this:
>
> /usr/sbin/exportfs -v
> /var/disks/sda <world> =
> (ro,async,wdelay,root_squash,no_subtree_check,anonuid=3D65534,anongid=3D6=
5 =
> 534)
> /var/disks/sdb <world> =
> (ro,async,wdelay,root_squash,no_subtree_check,anonuid=3D65534,anongid=3D6=
5 =
> 534)
> /var/disks/sdh <world> =
> (ro,async,wdelay,root_squash,no_subtree_check,anonuid=3D65534,anongid=3D6=
5 =
> 534)
> /var/disks/sdi <world> =
> (ro,async,wdelay,root_squash,no_subtree_check,anonuid=3D65534,anongid=3D6=
5 =
> 534)
On the server, mounting the web data file systems with "noatime" may =
help reduce the number of seeks on the disks.
Also, the "async" export option won't have any effect on reads.
Check your block device configuration as well. You may find that =
varying the RAID configuration, file system type (ext3 v. xfs) and =
stripe/chunk size could impact your server's performance. You might =
find that the deadline disk scheduler performs a little better than =
the default cfq scheduler.
It goes without saying that you should make sure your disk subsystem =
is healthy. I've found that, for example, SATA drives in hot-swap =
enclosures are sometimes affected by silent SATA transport errors =
that result in slow performance. Check dmesg carefully to ensure you =
are getting the highest possible speed settings. If only one of your =
drives is running significantly slower than the others, it will have =
a significant impact on the performance of a RAID group.
> on the client i mount them like this:
>
> lotta:/var/disks/sda on /var/disks/sda type nfs =
> (ro,hard,intr,proto=3Dtcp,rsize=3D32k,addr=3D217.213.5.44)
> lotta:/var/disks/sdb on /var/disks/sdb type nfs =
> (ro,hard,intr,proto=3Dtcp,rsize=3D32k,addr=3D217.213.5.44)
> lotta:/var/disks/sdh on /var/disks/sdh type nfs =
> (ro,hard,intr,proto=3Dtcp,rsize=3D32k,addr=3D217.213.5.44)
> lotta:/var/disks/sdi on /var/disks/sdi type nfs =
> (ro,hard,intr,proto=3Dtcp,rsize=3D32k,addr=3D217.213.5.44)
There are some client-side mount options that might also help. Using =
"nocto" and "actimeo=3D7200" could reduce synchonous NFS protocol =
overhead. I also notice a significant amount of readdirplus =
traffic. Readdirplus requests are fairly heavyweight, and in this =
scenario may be unneeded overhead. Your client might support the =
recently added "nordirplus" mount option, which could be helpful.
I wonder if "rsize=3D32k" is supported - you might want "rsize=3D32768" =
instead. Or better, let the client and server negotiate the maximum =
that each supports automatically by leaving this option off. You can =
check what options are in effect on each NFS mount point by looking =
in /proc/self/mountstats on the client.
Enabling jumbo frames between your NFS server and client will help. =
Depending on your NIC, though, it may introduce some instability =
(driver and hardware mileage may vary).
Since you currently have only one client, you might consider running =
the client and server back-to-back (ie replace any hub or switch with =
a simple cross-over link) to eliminate extra network overhead. =
Getting a high performance switch when you add more clients is key to =
making this configuration scale well -- a $99 special won't cut it.
> but when i then do the same dd again on the client i get
> disappointing 60-70Mbyte/s altogether. from a single disk i get
> ~25Mbytes/s on the client side.
25MB/s is fairly typical for Linux NFS servers.
> i played with some buffers /proc/sys/net/core/rmem_max and
> /proc/sys/net/core/rmem_default and increased them to 256M on the
> client.
You should consider similar network tuning on the server. Use a =
network benchmarking tool like iperf to assist.
> i was suspecting that the nfs server reads the files in too small
> chunks and tried to help it with
>
> for i in a h i ; do ( echo $((1024*6)) > /sys/block/sd$i/queue/ =
> read_ahead_kb ) ; done
>
> to get it to read in the files in one go.
Insufficient read-ahead on your server may be an issue here. Read =
traffic from the client often arrives at the server out of order, =
preventing the server from cleanly detecting sequential reads. I =
believe there was a recent change to the NFS server that addresses =
this issue.
> I would hope to at least double the speed.
IMO you can do that only by adding more clients.
> do
> you have a benchmark tool that can tell me the latency? i tried
> iozone and tried forcing it to only do read tests and did not get
> any helpfull error or output at all.
Use "iozone -a -i 1" to run read tests. You can narrow the test down =
to 2MB sequential reads if you want. Take a look at "iozone -h" =
output for more details.
> on the server:
>
> nfsstat
> Server rpc stats:
> calls badcalls badauth badclnt xdrcall
> 98188885 0 0 0 0
>
> Server nfs v3:
> null getattr setattr lookup access =
> readlink
> 5599 0% 318417 0% 160 0% 132643 0% 227130 0% =
> 0 0%
> read write create mkdir symlink mknod
> 97256921 99% 118313 0% 168 0% 0 0% 0 0% =
> 0 0%
> remove rmdir rename link readdir =
> readdirplus
> 162 0% 0 0% 0 0% 0 0% 0 0% =
> 105556 0%
> fsstat fsinfo pathconf commit
> 0 0% 1270 0% 0 0% 7153 0%
>
>
>
> cat /proc/net/rpc/nfsd
> rc 0 118803 98069945
> fh 0 0 0 0 0
> io 3253902194 38428672
> th 128 10156908 1462.848 365.212 302.100 252.204 311.632 187.508 =
> 142.708 142.132 198.168 648.640
> ra 256 97097262 0 0 0 0 0 0 0 0 0 64684
> net 98188985 16 98188854 5619
> rpc 98188885 0 0 0 0
> proc2 18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> proc3 22 5599 318417 160 132643 227130 0 97256921 118313 168 0 0 0 =
> 162 0 0 0 0 105556 0 1270 0 7153
> proc4 2 0 0
> proc4ops 40 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 =
> 0 0 0 0 0 0 0 0 0 0 0 0
--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
* Chuck Lever ([email protected]) [071025 20:25]:
> On Oct 25, 2007, at 9:10 AM, Andreas Schuldei wrote:
> >Hi!
> >
> >I need to tune a nfs server and client. on the server we have
> >several Tbyte of ~2Mbyte files and we need to transfer them read
> >only to the client. latency and throughput are crucial.
Because i have Tbytes of data but only a few Gbytes or RAM my
cache hits are rather unlikely. =
> >Right now i have only four disks in the server and i get 50Mbyte
> >out of each of them, simultaniously, for real world loads (random
> >reads across the disk, trying to minimizing the seeks by reading
> >the files in one go with
> >
> >for i in a b h i ; do ( find /var/disks/sd$i -type f | xargs -I=B0 dd if=
=3D=B0 bs=3D2M of=3D/dev/null status=3Dnoxfer =
> >2>/dev/null & ) ; done
> >
> >so with this (4*50 Mbyte/s) i should be able to saturate both
> >network cards.
note that this is my server's disk io performance.
> With a single client, you should not expect to get any better performance=
than by running the web service on the NFS =
> server. The advantage of using NFS under a web service is that you can t=
ransparently scale horizontally. When you add =
> a second or third web server that serves the same file set, you will see =
an effective increase in the size of the data =
> cache between your NFS server's disks and the web servers.
Not with terabyte of data and a distributed access pattern.
Certainly i will have some cache hits but not enough to be able
to serv considerable amounts out of RAM.
> But don't expect to get better data throughput over NFS than you see on y=
our local NFS server. =
That is exactly the point. on my server i get 4*50Mbytes =3D
200Mbyte/s out of the disks (with the above FOR loop around the
find and dd) and when i export on the same server the disks to an
nfs client i all of a sudden loose ~75% of the performance.
> If anything, the 10s =
> latency you see when the web server is on the same system with the disks =
is indicative of local file system =
> configuration issues.
how can i measure the latency on the local machine? i would be
very interested in seeing how it behaves latency wise.
> >on the server i start 128 nfs servers (RPCNFSDCOUNT=3D128) and export
> >the disks like this:
> >
> >/usr/sbin/exportfs -v
> >/var/disks/sda <world>(ro,async,wdelay,root_squash,no_subtree_check,anon=
uid=3D65534,anongid=3D65534)
> >/var/disks/sdb <world>(ro,async,wdelay,root_squash,no_subtree_check,anon=
uid=3D65534,anongid=3D65534)
> >/var/disks/sdh <world>(ro,async,wdelay,root_squash,no_subtree_check,anon=
uid=3D65534,anongid=3D65534)
> >/var/disks/sdi <world>(ro,async,wdelay,root_squash,no_subtree_check,anon=
uid=3D65534,anongid=3D65534)
> =
> On the server, mounting the web data file systems with "noatime" may help=
reduce the number of seeks on the disks.
yes, we do that already.
> >on the client i mount them like this:
> >
> >lotta:/var/disks/sda on /var/disks/sda type nfs (ro,hard,intr,proto=3Dtc=
p,rsize=3D32k,addr=3D217.213.5.44)
> >lotta:/var/disks/sdb on /var/disks/sdb type nfs (ro,hard,intr,proto=3Dtc=
p,rsize=3D32k,addr=3D217.213.5.44)
> >lotta:/var/disks/sdh on /var/disks/sdh type nfs (ro,hard,intr,proto=3Dtc=
p,rsize=3D32k,addr=3D217.213.5.44)
> >lotta:/var/disks/sdi on /var/disks/sdi type nfs (ro,hard,intr,proto=3Dtc=
p,rsize=3D32k,addr=3D217.213.5.44)
> =
> There are some client-side mount options that might also help. Using "no=
cto" and "actimeo=3D7200" could reduce =
> synchonous NFS protocol overhead. I also notice a significant amount of =
readdirplus traffic. Readdirplus requests are =
> fairly heavyweight, and in this scenario may be unneeded overhead. Your =
client might support the recently added =
> "nordirplus" mount option, which could be helpful.
> =
> I wonder if "rsize=3D32k" is supported - you might want "rsize=3D32768" i=
nstead.
i think that gave an effect. now i am in the 90-100Mbyte/s
ballpark and might hit the one-nic (1gbit) bottleneck.
> Or better, let the client and server =
> negotiate the maximum that each supports automatically by leaving this op=
tion off. You can check what options are in =
> effect on each NFS mount point by looking in /proc/self/mountstats on the=
client.
there it says now, after i specified rsize=3D2097152:
opts: rw,vers=3D3,rsize=3D1048576,wsize=3D1048576,acregmin=3D3,ac=
regmax=3D60,acdirmin=3D30,acdirmax=3D60,hard,intr,nolock,proto=3Dtcp,timeo=
=3D600,retrans=3D2,sec=3Dsys
i am surprised that it did not protest when it could not parse
the "k". note that it it only took 1M chunks. how come?
> Enabling jumbo frames between your NFS server and client will help. Depe=
nding on your NIC, though, it may introduce =
> some instability (driver and hardware mileage may vary).
i will test that and bonding two nicks.
> Insufficient read-ahead on your server may be an issue here. Read traffi=
c from the client often arrives at the server =
> out of order, preventing the server from cleanly detecting sequential rea=
ds. I believe there was a recent change to =
> the NFS server that addresses this issue.
when did that go in? do i need to activate that somehow?
how can i measure the latency on a loaded server? both locally
and over nfs?
/andreas
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs
At 03:34 PM 10/25/2007, Andreas Schuldei wrote:
>> But don't expect to get better data throughput over NFS than you see
>on your local NFS server.
>
>That is exactly the point. on my server i get 4*50Mbytes =
>200Mbyte/s out of the disks (with the above FOR loop around the
>find and dd) and when i export on the same server the disks to an
>nfs client i all of a sudden loose ~75% of the performance.
Andreas - you mentioned earlier in the thread that you're using 2.6.22
as the server kernel. You might consider moving it forward to 2.6.23,
as there is a significant readahead improvement/fix there. When you
run a local read, all readahead is perfect. However on older kernels
serving NFS, the reordering of reads due to Linux's multithreaded nfsd's
triggers several issues which spook the readahead into truly awful
behavior.
Fengguang Wu's changes are merged in 2.6.23, in testing with NFS/RDMA
we now see the same performance over the wire as locally, for any
number of readers.
My one caveat to this is that your files are relatively small, so I would
not expect a truly dramatic speedup from it. It's certainly worth trying,
there are other improvements in 2.6.23.
Do you have Infiniband or iWARP adapters in your environment? If you're
worried about bandwidth and application startup latency, using it could
certainly help. You'd need to run a 2.6.24-rc1 client though, and some
assembly is still required to use the server (it's not yet upstream).
Tom.
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs