Dear Hedok, Darren and others,

I promise to provide some numbers as soon as I have some time. But a few
thougths on your observations:

Neither a simple file transfer nor a benchmark on SPM segmentation are a
good stress test for filesystem performance in SPM, because
- file transfer only tests _either_ read _or_ write performance
- segmentation loads data once and does a lengthy computation before
   writing once.

A more realistic test is e.g. realignment, writing normalised images,
smoothing or running a fMRI first level stats estimation, where you have
lots of data to read, less data to write (and to different files than the
originals).

On Fri, 6 Apr 2007, Hedok Lee wrote:

> Dear Darren and SPMers.
>
> There are only two computers in my set up, and there is no NAS. One computer
> for running SPM analyses and another one for storage. I've been considering
> NAS, but it was difficult to find the ones which support NFS. NFS and FTP
> are the only servers that I'm currently running.
>
> As for performance improvements, exporting NFS as "async" has improved
> transfer rate by 300%. Also, people use "Jumbo" frame by changing packet
> size from 1024 to 8192. I ran some testing to see the differences in
> performance under different conditions.
>
> This link may help you about "async"
> http://www.mythtv.org/wiki/index.php/Optimizing_Performance
>
> I'm no IT guru, but try to provide some information that may be useful.
>
> Computers
> Intel P4(3.0Gx2) Non-RAID SCSI 10000rpm built-in gigabit card by DELL.
> running RH4 WS with ext3 partition
> Intel P3(1.0x2) Non-RAID IDE 7200rpm PCI gigabit card. running CentOS4
> with ext3 partition
> Router: Linksys 5ports gigabit with full duplex support.
> Note: Not sure if these cards support jumbo frame.
>
> Files to transfer
> 167MB(1297 DICOM files by Siemens) x5=835MB
>
> Total time it took to copy 835MB from P4->P3
> 216sec(3.9MB/sec) default setting
> 74sec(11.3MB/sec) async turned on
> 63sec(13MB/sec) async+Jumbo frame turned on
>
> Total time it took to copy 835MB from P3->P4
> 29sec(29MB/sec) default setting
> 30sec(28MB/sec) async turned on
> 36sec(23MB/sec) async+Jumbo frame turned on
> Note: I have no idea why it's so much faster when I transfer P3->P4. Perhaps
> it has to with writing speed in SCSI?
>
> SPM test
> Default SPM5 segmentation of 3 high resolution anatomical images, SPGR
> (1.0x1.0x1.5)
> Total time to complete the segmentation
> Local storage (SCSI): 25minutes
> Remote storage (NFS): 25minutes
>
> Conclusion:
> I'm beginning to think that SPM over NFS may be ok as total time it took to
> complete segmentation was the same.
>
> I'm still curious if I can improve the performance.
>
> Hedok
>
>
> [log in to unmask] wrote:
>> Dear Hedok, Volkmar,
>>
>> I'd be interested in hearing about (on or off list) what file systems
>> (nfs, samba, etc) you are
>> using to enable higher performance data transfers on linux systems for mri
>> data analysis. Are you
>> using NAS-type of data storage?
>>
>> thanks
>> darren
>>
>> ==============Original message text===============
>> On Thu, 05 Apr 2007 1:26:23 pm CDT Hedok Lee wrote:
>>
>> Dear Volkmar and SPMers
>>
>> Regarding this post,
>>
>> http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0704&L=SPM&P=R3600&I=-3
>>
>> > I've just investigated this because I am doing a transition of our
>> > network filesystem away from
>> >
>> NFS and had to resolve some performance issues as well.
>>
>> As our data continue to grow, I was thinking about having a storage server
>> and mount the data
>> through NFS. This way, I can keep adding more hard drives onto a storage
>> server if necessary. I
>> was wondering if there is any particular reason why you decided to move
>> away from NFS. If you have a better solution to manage storage server,
>> please let me know. All of our machines are linux
>> based, and they are connected through gigabit but my experience with NFS
>> has been pretty slow
>> ~ 15MB/sec. I tried jumbo frame and etc, but had not luck to reach
>> ~ ~40MB/sec.
>>
>> This is probably too slow to carry out SPM analysis over NFS isn't it?
>>
>> With regards,
>>
>> Hedok
>>
>>
>>
>>
>> ----- Original Message -----
>> From: Volkmar Glauche <[log in to unmask]>
>> Date: Thursday, April 5, 2007 3:42 am
>> Subject: Re: SPM5 slower with 4D files
>> To: [log in to unmask]
>>
>>
>> > Hi Jerome,
>> >
>> > yes, SPM will probably slower on 4D files than on 3D files. Also it will
>> > be slower on .nii files than on .hdr/.img pairs. The reason is, that for
>> > combined header and data files (as in 3D .nii and 4D cases) SPM needs
>> > something like random-write-access to your file to store small bits of
>> > information. This is more expensive in terms of I/O time (especially on
>> > large disks/RAID systems/network file storage) than storing 3D data with
>> > separate .hdr/.img files. However, storing one large file is more
>> > efficient in terms of disk usage (if you have partitioned/formatted your
>> > disk to efficiently store large files). In short, there is no free
>> > lunch, and you have to decide what is more important to you: saving time
>> > or saving disk space...
>> > I've just investigated this because I am doing a transition of our
>> > network filesystem away from NFS and had to resolve some performance
>> > issues as well.
>> >
>> > Volkmar
>> >
>> > On Wed, 4 Apr 2007, Paul Macey wrote:
>> >
>> >
>> > > Hi Jérôme
>> > >
>> > > I just wrote something to split 4D to 3D; run the attached
>> > >
>> > script
>> >
>> > > (cspm_im_4Dto3D, from SPM5 only) and it will prompt for a 4D
>> > >
>> > file to split.
>> >
>> > > You can also use it in batch mode with a couple more options
>> > >
>> > (see help). The
>> >
>> > > 3D files should be the same as in the 4D file, so I don't think
>> > >
>> > you will
>> >
>> > > loose anything by converting.
>> > >
>> > > Best wishes,
>> > > Paul
>> > >
>> > > Jérôme Redouté wrote:
>> > >
>> > > > Dear SPMers,
>> > > > My parameters estimation process is extremely slow. Could
>> > > >
>> > this slowness
>> >
>> > > > be due to using a 4D-Niftii dataset?
>> > > > In order to check that, I would like to convert my sw_xxxx.nii
>> > > >
>> > 4D files
>> >
>> > > > into a set of 3D files.
>> > > > My questions are:
>> > > > -SPM has a function to convert 3D -> 4D , but is the reverse
>> > > >
>> > function>> available? Where can I find such a converter?
>> >
>> > > > -These images are preprocessed (realignment, normalisation,
>> > > >
>> > smoothing), do
>> >
>> > > > I risk to lose these parameters by converting to 4D to 3D?
>> > > > Thanks for your help
>> > > > Jerome
>> > > >
>> > > >
>> > --
>> > Volkmar Glauche
>> > -
>> > Department of Neurology [log in to unmask]
>> > Universitaetsklinikum Freiburg Phone 49(0)761-270-5331
>> > Breisacher Str. 64 Fax 49(0)761-270-5416
>> > 79106 Freiburg
>> > http://fbi.uniklinik-freiburg.de/===========End of original message
>> > text===========
>> >
>>
>>
>>
>> --------------------------------
>> Darren R. Gitelman, M.D.
>> Department of Neurology
>> 710 N. Lake Shore Dr. #1122
>> Chicago, IL 60611
>> Voice: (312) 908-8614
>> Fax: (312) 908-5073
>> --------------------------------
>>
>>
>>
>
>

--
Volkmar Glauche
-
Department of Neurology [log in to unmask]
Universitaetsklinikum Freiburg Phone 49(0)761-270-5331
Breisacher Str. 64 Fax 49(0)761-270-5416
79106 Freiburg http://fbi.uniklinik-freiburg.de/