[Lustre-devel] Lustre and cross-platform portability
Andreas Dilger
As a prerequisite to this, EMC is working to clean up the Lustre client code to better match the kernel coding style, and one of the anticipated major obstacles to upstream kernel submission is the heavy use of code abstraction via libcfs for portability to other operating systems (most notably MacOS and WinNT, but also for liblustre, and potentially *BSD).
I have no information that the WinNT project will ever be released by Oracle, and as yet there has not been any code released from the MacOS port, so the libcfs portability layer is potentially exacting a high cost in code maintenance and complexity (CLIO being a prime example) for no apparent benefit. Similarly, the liblustre client needs a portability layer for userspace, and suffers from the same apparent lack of interest or users.
I'd like to get some feedback from the Lustre community about removing the libcfs abstraction entirely, or possibly restructuring it to look like the Linux kernel API, and having the other platforms code against it as a Linux portability layer, like ZFS on Linux uses the Solaris Portability Layer (SPL) to avoid changing the core ZFS code. A related topic is whether it would be better to replace all cfs_* functions with standard Linux kernel functions en-masse, or migrate away from cfs_* functions slowly?
Also, we're planning on deprecating the liblustre client code, due to lack of interest/usage. The current code is in disrepair, and we've been keeping it around for years without any benefit, and while I was one of the strongest advocates for keeping it in our back pocket in case of future needs, I don't see that materializing in the future.
The liblustre code would be left in the tree for now, in case someone from the community is interested to get it working and maintain it, and it may be updated on a best effort basis. If nobody steps forward to do this work, the liblustre code would be deleted from the development branch in a year or so.
Unfortunately, after starting this thread, I may not be able to reply to questions in a timely manner due to vacation. I look forward to a thread that concludes with unanimous agreement from all parties. :-)
Cheers, Andreas
--
Andreas Dilger Whamcloud, Inc.
Principal Lustre Engineer http://www.whamcloud.com/
_______________________________________________
Lustre-devel mailing list
Lustre...@lists.lustre.org
http://lists.lustre.org/mailman/listinfo/lustre-devel
Ken Hornstein
>by Oracle, and as yet there has not been any code released from the
>MacOS port, so the libcfs portability layer is potentially exacting
>a high cost in code maintenance and complexity (CLIO being a prime
>example) for no apparent benefit. Similarly, the liblustre client needs
>a portability layer for userspace, and suffers from the same apparent
>lack of interest or users.
In terms of the MacOS X port, I don't think that's quite fair ...
the code I did is available and anyone can download it. It was
functional in a very basic way but needed some additonal love.
Okay, I haven't rolled that stuff into the Whamcloud release ...
what happened there was when there was all the uncertainty with
Oracle & Lustre development I lost momentum and got caught up in
other things. I've talked with the guys at Whamcloud about bringing
at least the portability changes over, and that's all been on me;
it's certainly on my list to work on.
I can say that at least for MacOS X, there has been interest; I can't
speak for the amount of interest, and there's a bit of a chicken and
egg problem ... people don't plan their Lustre use around MacOS X
clients because there isn't one that works well, and people don't put
work into it because there isn't people who plan their Lustre use
around it.
>I'd like to get some feedback from the Lustre community about removing
>the libcfs abstraction entirely, or possibly restructuring it to look
>like the Linux kernel API, and having the other platforms code against
>it as a Linux portability layer, like ZFS on Linux uses the Solaris
>Portability Layer (SPL) to avoid changing the core ZFS code. A related
>topic is whether it would be better to replace all cfs_* functions with
>standard Linux kernel functions en-masse, or migrate away from cfs_*
>functions slowly?
The only thing I can think of is that if this is done, then officially
Lustre is going to be a Linux-only filesystem. I understand there are
real costs to maintaining the cfs layer, and I can't speak to whether or
not the loss would be worth the gains.
--Ken
Andreas Dilger
>> I have no information that the WinNT project will ever be released
>> by Oracle, and as yet there has not been any code released from the
>> MacOS port, so the libcfs portability layer is potentially exacting
>> a high cost in code maintenance and complexity (CLIO being a prime
>> example) for no apparent benefit. Similarly, the liblustre client
>> needs a portability layer for userspace, and suffers from the same
>> apparent lack of interest or users.
>
> In terms of the MacOS X port, I don't think that's quite fair ...
> the code I did is available and anyone can download it.
Ken, my apologies for this misstatement. I guess that my faulty memory is to blame for the fact that I didn't recall the MacOS code was made publicly available for download.
> It was
> functional in a very basic way but needed some additonal love.
> Okay, I haven't rolled that stuff into the Whamcloud release ...
I don't think I've ever seen patches sent from you to either Oracle or Whamcloud, and unfortunately nobody on our side has had the bandwidth or user demand/funding to be pulling such changes either.
> what happened there was when there was all the uncertainty with
> Oracle & Lustre development I lost momentum and got caught up in
> other things. I've talked with the guys at Whamcloud about bringing
> at least the portability changes over, and that's all been on me;
> it's certainly on my list to work on.
>
> I can say that at least for MacOS X, there has been interest; I can't
> speak for the amount of interest, and there's a bit of a chicken and
> egg problem ... people don't plan their Lustre use around MacOS X
> clients because there isn't one that works well, and people don't put
> work into it because there isn't people who plan their Lustre use
> around it.
Another alternative that has been proposed is FUSE for MacOS/Solaris. I'm not sure if FUSE exists for Windows or not. Of course, this is only useful for casual access, and not for high performance usage.
>> I'd like to get some feedback from the Lustre community about removing
>> the libcfs abstraction entirely, or possibly restructuring it to look
>> like the Linux kernel API, and having the other platforms code against
>> it as a Linux portability layer, like ZFS on Linux uses the Solaris
>> Portability Layer (SPL) to avoid changing the core ZFS code. A related
>> topic is whether it would be better to replace all cfs_* functions with
>> standard Linux kernel functions en-masse, or migrate away from cfs_*
>> functions slowly?
>
> The only thing I can think of is that if this is done, then officially
> Lustre is going to be a Linux-only filesystem. I understand there are
> real costs to maintaining the cfs layer, and I can't speak to whether or
> not the loss would be worth the gains.
This isn't strictly correct. It would be possible to change the libcfs portability layer to export the same API as the Linux kernel to MacOS and Windows. This would simplify getting the client into the Linux kernel, but still allow a native client on MacOS.
I'm fine with keeping the abstraction layer if there is legitimate use for it. Having the MacOS client in the Lustre tree in a state where it can at least be built would allow us to make changes to the code (e.g. function renaming) with some chance of it not being totally broken immediately.
Cheers, Andreas
--
Andreas Dilger Whamcloud, Inc.
Principal Lustre Engineer http://www.whamcloud.com/
Joshua Walgenbach
Hash: SHA1
On 03/15/2012 03:39 PM, Andreas Dilger wrote:
> Another alternative that has been proposed is FUSE for
> MacOS/Solaris. I'm not sure if FUSE exists for Windows or not. Of
> course, this is only useful for casual access, and not for high
> performance usage.
>
I think the FUSE for Windows project is defunct, but there is a user
space filesystem project for Windows called Dokan
(http://dokan-dev.net/en/) that had a release a little more than a
year ago.
For my part, a Lustre client on Windows or OS X would be used mostly
for visualization of data, rather than being computed against so a
slower user space implementation would be more than sufficient. There
are a few applications that are using an SMB exported Lustre
filesystem for data collection, but those applications are similarly
low bandwidth. It would be nice to remove the SMB server in the middle.
- -Josh
- --
Joshua Walgenbach, High Performance File Systems, Indiana University
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
iEYEARECAAYFAk9iSFEACgkQcqyJPuRTYp8dRgCePbaKA2P1Km80KBJM/YG6Nky1
8pgAn3RQW1i5omkhJfFZOG+0yhtXKJFB
=08vB
-----END PGP SIGNATURE-----
Alexey Lyashkov
but using a full client will able to use a striping information and accept a data from a several OST's
with comparing a get a single data stream from a SMB.
Also fuse client will able to run on any OS have a FUSE porting that is any BSD, OpenSolaris, MacOS, in additional to the windows.
That is easy way to maintain a single client for many OS.
On Mar 16, 2012, at 14:11, Gregory Matthews wrote:
> On 15/03/12 19:51, Joshua Walgenbach wrote:
>> For my part, a Lustre client on Windows or OS X would be used mostly
>> for visualization of data, rather than being computed against so a
>> slower user space implementation would be more than sufficient.
>> There are a few applications that are using an SMB exported Lustre
>> filesystem for data collection, but those applications are similarly
>> low bandwidth. It would be nice to remove the SMB server in the
>> middle.
>
> why bother having a windows client if you lose the performance? We have
> windows based detectors and proprietary windows based analysis software
> that would definitely benefit from higher performance access to lustre
> file systems but replacing existing CIFS servers for no gain seems a bit
> pointless.
>
> I understand the problems that a native MS client would have, without
> buy-in from Microsoft it would be a nightmare to keep the MS tree and
> the Linux tree in sync. One possible future workaround proposed by
> Whamcloud at last years LUG is pNFS.
>
> GREG
>
>>
>> - -Josh
>
>
> --
> Greg Matthews 01235 778658
> Senior Computer Systems Administrator
> Diamond Light Source, Oxfordshire, UK
>
> --
> This e-mail and any attachments may contain confidential, copyright and or privileged material, and are for the use of the intended addressee only. If you are not the intended addressee or an authorised recipient of the addressee please notify us of receipt by returning the e-mail and do not use, copy, retain, distribute or disclose the information in or attached to the e-mail.
> Any opinions expressed within this e-mail are those of the individual and not necessarily of Diamond Light Source Ltd. Diamond Light Source Ltd. cannot guarantee that this e-mail or any attachments are free from viruses and we cannot accept liability for any damage which you may sustain as a result of software viruses which may be transmitted in or with the message.
> Diamond Light Source Limited (company no. 4375679). Registered in England and Wales with its registered office at Diamond House, Harwell Science and Innovation Campus, Didcot, Oxfordshire, OX11 0DE, United Kingdom
Ken Hornstein
>is to blame for the fact that I didn't recall the MacOS code was made
>publicly available for download.
No problem. Back when I gave the talk at LUG the source wasn't available
yet due to issues here, but we got that worked out and I was pushing
my changes to a publically available Oracle git repo. I did send out
email to everyone about that, but I'm sure it was easy to miss.
>I don't think I've ever seen patches sent from you to either Oracle or
>Whamcloud, and unfortunately nobody on our side has had the bandwidth or
>user demand/funding to be pulling such changes either.
Well, I did actually submit patches to Oracle to start the process of
working out at least the portability issues, but I believe that was
when Oracle started to implode the Lustre group so things sort of
stalled. I'll take 75% of the blame for that if we assign 25% to
Larry Ellison :-)
>This isn't strictly correct. It would be possible to change the libcfs
>portability layer to export the same API as the Linux kernel to MacOS
>and Windows. This would simplify getting the client into the Linux
>kernel, but still allow a native client on MacOS.
Well ... that shifts the burden to cross-platform people basically having
to re-implment the Linux kernel. For some things, that's possible without
too much pain. For other things, it's not.
--Ken
Ken Hornstein
>any BSD, OpenSolaris, MacOS, in additional to the windows. That is
>easy way to maintain a single client for many OS.
It is, unfortunately, not quite that simple.
I can't claim to be a FUSE expert, but I've been paying attention
to it on other platforms. From what I can tell, FUSE works great
on Linux, but on other platforms the support is iffy. Also, it's
not quite implemented the same on other operating systems as it is
on Linux, making porting a Linux FUSE module to other platforms not
trivial; from what I've seen, this is due to the Linux filesystem
interface versus the vnode interface used by every Unix except Linux
(and this is part of what makes Lustre hard to port).
I guess what I'm saying is that don't fall into the underwear gnomes
trap of thinking:
1) Get liblustre working with FUSE
2) ???
3) Lustre client everywhere!
It might make it easier, but I doubt it will make it easy.
--Ken
Ward, Lee
On Mar 14, 2012, at 6:31 PM, Andreas Dilger wrote:
> Whamcloud and EMC are jointly investigating how to be able to contribute the Lustre client code into the upstream Linux kernel.
>
> As a prerequisite to this, EMC is working to clean up the Lustre client code to better match the kernel coding style, and one of the anticipated major obstacles to upstream kernel submission is the heavy use of code abstraction via libcfs for portability to other operating systems (most notably MacOS and WinNT, but also for liblustre, and potentially *BSD).
>
> I have no information that the WinNT project will ever be released by Oracle, and as yet there has not been any code released from the MacOS port, so the libcfs portability layer is potentially exacting a high cost in code maintenance and complexity (CLIO being a prime example) for no apparent benefit. Similarly, the liblustre client needs a portability layer for userspace, and suffers from the same apparent lack of interest or users.
>
> I'd like to get some feedback from the Lustre community about removing the libcfs abstraction entirely, or possibly restructuring it to look like the Linux kernel API, and having the other platforms code against it as a Linux portability layer, like ZFS on Linux uses the Solaris Portability Layer (SPL) to avoid changing the core ZFS code. A related topic is whether it would be better to replace all cfs_* functions with standard Linux kernel functions en-masse, or migrate away from cfs_* functions slowly?
>
> Also, we're planning on deprecating the liblustre client code, due to lack of interest/usage. The current code is in disrepair, and we've been keeping it around for years without any benefit, and while I was one of the strongest advocates for keeping it in our back pocket in case of future needs, I don't see that materializing in the future.
Relatively true I think, unfortunately. In the past, we provided direct support of Lustre for processes running on light-weight kernels via liblustre. I am aware that DOE/NNSA (more than just Sandia) believes light-weight kernels are the future but it seems that it may be quite a while, yet, before we would be forced into that choice. I don't see NNSA having some sort of heart-attack over your obfuscating liblustre then.
--Lee
Alexey Lyashkov
I (with two friends) write a FUSE client for a FreeBSD in past.
Most issues related to a porting libcfs/libsysio to FreeBSD, but anyway patch isn''t huge ~200kb.
I have plan to send that patch in near time, as 2.3 landing window open.
Gregory Matthews
> for visualization of data, rather than being computed against so a
> slower user space implementation would be more than sufficient.
> There are a few applications that are using an SMB exported Lustre
> filesystem for data collection, but those applications are similarly
> low bandwidth. It would be nice to remove the SMB server in the
> middle.
why bother having a windows client if you lose the performance? We have
windows based detectors and proprietary windows based analysis software
that would definitely benefit from higher performance access to lustre
file systems but replacing existing CIFS servers for no gain seems a bit
pointless.
I understand the problems that a native MS client would have, without
buy-in from Microsoft it would be a nightmare to keep the MS tree and
the Linux tree in sync. One possible future workaround proposed by
Whamcloud at last years LUG is pNFS.
GREG
>
> - -Josh
--
Greg Matthews 01235 778658
Senior Computer Systems Administrator
Diamond Light Source, Oxfordshire, UK
--
This e-mail and any attachments may contain confidential, copyright and or privileged material, and are for the use of the intended addressee only. If you are not the intended addressee or an authorised recipient of the addressee please notify us of receipt by returning the e-mail and do not use, copy, retain, distribute or disclose the information in or attached to the e-mail.
Any opinions expressed within this e-mail are those of the individual and not necessarily of Diamond Light Source Ltd.
Diamond Light Source Ltd. cannot guarantee that this e-mail or any attachments are free from viruses and we cannot accept liability for any damage which you may sustain as a result of software viruses which may be transmitted in or with the message.
Diamond Light Source Limited (company no. 4375679). Registered in England and Wales with its registered office at Diamond House, Harwell Science and Innovation Campus, Didcot, Oxfordshire, OX11 0DE, United Kingdom
Todd, Allen
On Friday, March 16, 2012 6:11 AM, Gregory Matthews wrote:
> why bother having a windows client if you lose the performance? We have
> windows based detectors and proprietary windows based analysis software
> that would definitely benefit from higher performance access to lustre
> file systems but replacing existing CIFS servers for no gain seems a bit
> pointless.
>From the perspective of my firm, the benefit of a windows lustre client (native or fuse-based), even one that performs at 10% or 20% of the linux client, is the improved scalability that it offers. Our current solution uses 4 samba gateways per hundred windows servers to achieve acceptable bandwidth, but that bandwidth can be cut to an unacceptable trickle if a large wave of native linux clients simultaneously accesses the filesystem.
We have repeatedly asked our Microsoft HPC contacts to intervene to either fund a native windows client or to get oracle to release the existing one, since scalable storage is a big hole in the Microsoft HPC toolkit. Obviously, nothing has come out of that. It seems Microsoft is banking on pNFS, eventually working in this space.
Allen Todd
________________________________
IMPORTANT: The information contained in this email and/or its attachments is confidential. If you are not the intended recipient, please notify the sender immediately by reply and immediately delete this message and all its attachments. Any review, use, reproduction, disclosure or dissemination of this message or any attachment by an unintended recipient is strictly prohibited. Neither this message nor any attachment is intended as or should be construed as an offer, solicitation or recommendation to buy or sell any security or other financial instrument. Neither the sender, his or her employer nor any of their respective affiliates makes any warranties as to the completeness or accuracy of any of the information contained herein or that this message or any of its attachments is free of viruses.
Nathan Rutman
On Mar 16, 2012, at 8:06 AM, Todd, Allen wrote:
>
> On Friday, March 16, 2012 6:11 AM, Gregory Matthews wrote:
>> why bother having a windows client if you lose the performance? We have
>> windows based detectors and proprietary windows based analysis software
>> that would definitely benefit from higher performance access to lustre
>> file systems but replacing existing CIFS servers for no gain seems a bit
>> pointless.
1. Re-exporting Lustre via CIFS or NFS isn't scalable to very large numbers of Windows clients
2. Re-exporting Lustre via CIFS or NFS can have coherency problems when multiple re-exporters are involved
3. Wide-striped file access would likely have greater performance on a native client than via a re-exported single pipe, for some value of "wide".