Skip to contentSkip to conversation

Would SSI clustering extensions be of interest to kernel community?

DISCUSSION Constantine Gavrilov 6 participants 10 messages Oct 16, 2006 – Oct 23, 2006 1 message not in the archive

lore
Share & tools
lore
#0 lore raw

I have implemented SSI (single system image) clustering extensions to Linux kernel in the form of a loadable module.

It roughly mimics OpenMosix model of deputy/remote split (migrated processes leave a stub on the node where they were born and depend on the "home" node for IO).

The implementation shares no code with Mosix/Open Mosix (was written from scratch), is much smaller, and is easily portable to multiple architectures.

We are considering publication of this code and forming an open source project around it.

I have two questions to the community:

1) Is community interested in using this code? Do users require SSI product in the era when everybody is talking about partitioning of machines and not clustering?
2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

Please CC me and the list when replying.

Signature · 8 lines
----------------------------------------
Constantine Gavrilov
Kernel Developer
Qlusters Software Ltd
1 Azrieli Center, Tel-Aviv
Phone: +972-3-6081977
Fax:   +972-3-6081841
----------------------------------------
#1 lore raw

Ar Llu, 2006-10-16 am 14:49 +0200, ysgrifennodd Constantine Gavrilov:

Quoted 4 lines

2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

If they are doing SSI well and do not need core kernel changes then yes they sound very interesting to me. Historically the big concern has always been that things like this muck up the kernel core which affects the other 99.99999% of users who don't want SSI clustering.

Alan

#5 lore raw
In reply to Constantine Gavrilov
Quoted 3 lines · Constantine Gavrilov

1) Is community interested in using this code? Do users require SSI product in the era when everybody is talking about partitioning of machines and not clustering?

Why not, I certainly like the idea of partitioning my machine with lots of XEN VMs and then joining all their power with SSI!

(couldn't resist it sorry)

#2 lore raw

Re: Would SSI clustering extensions be of interest to kernelcommunity?

In reply to Alan Cox
Quoted 7 lines · Alan Cox

Ar Llu, 2006-10-16 am 14:49 +0200, ysgrifennodd Constantine Gavrilov:

Quoted 3 lines

2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

Quoted 10 lines

If they are doing SSI well and do not need core kernel changes then yes they sound very interesting to me. Historically the big concern has always been that things like this muck up the kernel core which affects the other 99.99999% of users who don't want SSI clustering.

Alan

SSI intrudes kernel in two places: a) IO system calls, b ) page fault code for shared memory pages.

a) IO system calls are "packed" and forwarded to the "home" node, where original syscall code is executed.
b) A hook is inserted into page fault code that brings shared memory pages from other nodes when necessary.

Apart from these two hooks, SSI code is a "standalone" kernel API add-on ("add", not "change").

Currently, we can do both "intrusions" from the kernel module. I assume that if we submit code, you will require a kernel patch that explicitly calls our hooks.

Also, continuous SSI in-kernel support may require SSI changes in the following cases: a) new fields in task struct that reflect process state (may affect task migration), b) changes in the page fault mechanism (may effect SSI shared memory code that brings and invalidates pages), c) addition of new system calls (may require implementation of SSI suspport for them).

Signature · 8 lines
----------------------------------------
Constantine Gavrilov
Kernel Developer
Qlusters Software Ltd
1 Azrieli Center, Tel-Aviv
Phone: +972-3-6081977
Fax:   +972-3-6081841
----------------------------------------
#4 lore raw

Re: Would SSI clustering extensions be of interest to kernelcommunity?

In reply to a message not in the archive

Ar Llu, 2006-10-16 am 16:07 +0200, ysgrifennodd Constantine Gavrilov:

Quoted 8 lines

SSI intrudes kernel in two places: a) IO system calls, b ) page fault code for shared memory pages.

a) IO system calls are "packed" and forwarded to the "home" node, where original syscall code is executed.
b) A hook is inserted into page fault code that brings shared memory pages from other nodes when necessary.

Quoted 6 lines

Apart from these two hooks, SSI code is a "standalone" kernel API add-on ("add", not "change").

Currently, we can do both "intrusions" from the kernel module. I assume that if we submit code, you will require a kernel patch that explicitly calls our hooks.

Yep. Thats probably the most critical single thing to review.

Quoted 7 lines

Also, continuous SSI in-kernel support may require SSI changes in the following cases: a) new fields in task struct that reflect process state (may affect task migration), b) changes in the page fault mechanism (may effect SSI shared memory code that brings and invalidates pages), c) addition of new system calls (may require implementation of SSI suspport for them).

SSI changes triggered from core changes are fairly expected I think because you need to serialize new objects.

#6 lore raw
In reply to Constantine Gavrilov
Quoted 11 lines · Constantine Gavrilov

I have implemented SSI (single system image) clustering extensions to Linux kernel in the form of a loadable module.

It roughly mimics OpenMosix model of deputy/remote split (migrated processes leave a stub on the node where they were born and depend on the "home" node for IO).

The implementation shares no code with Mosix/Open Mosix (was written from scratch), is much smaller, and is easily portable to multiple architectures.

Quoted 8 lines

We are considering publication of this code and forming an open source project around it.

I have two questions to the community:

1) Is community interested in using this code? Do users require SSI product in the era when everybody is talking about partitioning of machines and not clustering?

Some users require SSI clustering and some just like playing with it. In any case, more options than those available currently can only be good :)

Quoted 5 lines

2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

I'm sure there's interrest in at least seeing it.
You should consider cleaning up your code according to Documentation/CodingStyle first though (if it doesn't already follow it) or your first batch of feedback is probably just going to be a bunch of style cleanup requests ;)

Signature · 3 lines
Jesper Juhl <[email protected]>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html
#7 lore raw
In reply to Constantine Gavrilov
Quoted 6 lines · Constantine Gavrilov

I have implemented SSI (single system image) clustering extensions to Linux kernel in the form of a loadable module.

It roughly mimics OpenMosix model of deputy/remote split (migrated processes leave a stub on the node where they were born and depend on the "home" node for IO).

The implementation shares no code with Mosix/Open Mosix (was written from scratch), is much smaller, and is easily portable to multiple architectures.

Quoted 9 lines

We are considering publication of this code and forming an open source project around it.

I have two questions to the community:

1) Is community interested in using this code? Do users require SSI product in the era when everybody is talking about partitioning of machines and not clustering?
2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

Please CC me and the list when replying.

I am interested in seeing the changes. I am right now working on getting parts of OpenSSI (www.openssi.org) changes merged upstream. Bruce Walker of the OpenSSI project have a design of implementing cluster wide procs. The same doc can be found on www.openssi.org website. The paper talks about how to implement cluster wide proccess model without requiring home/deputy concept. But yes it require some core kernel changes. But should be Conditionally enabled like selinux. So overhead for non cluster users should be nill.

Regarding my work you can see the status here
http://git.openssi.org/~kvaneesh/gitweb.cgi?p=ci-to-linus.git;a=summary

It only gets the ICS changes. That means it introduce a transport independent kernel cluster framework. Right now it supports two interconnect IPV4 and infiniband verbs.

I am planning on taking the CFS changes. That should bring in clusterwide shared memory too. The way it was done in OpenSSI was to hook a new nopage() function for CFS so that when we page fault, we bring the pages from other node.So i am not sure whether one need a VM hook for getting clusterwide shared memory. But without seeing the code i am clueless.

-aneesh

#8 lore raw

Please see inline...

Quoted 6 lines · Aneesh Kumar K.V

I am interested in seeing the changes. I am right now working on getting parts of OpenSSI (www.openssi.org) changes merged upstream. Bruce Walker of the OpenSSI project have a design of implementing cluster wide procs. The same doc can be found on www.openssi.org website. The paper talks about how to implement cluster wide proccess model without requiring home/deputy concept. But yes it require some core kernel changes. But should be Conditionally enabled like selinux. So overhead for non cluster users should be nill.

I am personally not interested in making intrusive kernel changes even if it yields in true "single-system image". I want very small changes (preferrably none).

Quoted 5 lines

Regarding my work you can see the status here
http://git.openssi.org/~kvaneesh/gitweb.cgi?p=ci-to-linus.git;a=summary

It only gets the ICS changes. That means it introduce a transport independent kernel cluster framework. Right now it supports two interconnect IPV4 and infiniband verbs.

We also have transport abstraction layer and transport plugins for TCP/IP, SDP (Infiniband and possibly others), and SCI (Dolphin).

Quoted 4 lines

I am planning on taking the CFS changes. That should bring in clusterwide shared memory too. The way it was done in OpenSSI was to hook a new nopage() function for CFS so that when we page fault, we bring the pages from other node.So i am not sure whether one need a VM hook for getting clusterwide shared memory. But without seeing the code i am clueless.

Nopage will be called if there is no pte. That means, with just nopage you cannot implement RO-RW transition. If you use nopage only, you cannot have multiple readers, because you cannot invalidate all other readers if one reader goes read-write. Thus nopage allows single reader or single writer whle the page fault hook allows multiple readers and single writer.

Signature · 8 lines
----------------------------------------
Constantine Gavrilov
Kernel Developer
Qlusters Software Ltd
1 Azrieli Center, Tel-Aviv
Phone: +972-3-6081977
Fax:   +972-3-6081841
----------------------------------------
#9 lore raw
Quoted 14 lines · Constantine Gavrilov

Please see inline...

Quoted 6 lines · Aneesh Kumar K.V

I am interested in seeing the changes. I am right now working on getting parts of OpenSSI (www.openssi.org) changes merged upstream. Bruce Walker of the OpenSSI project have a design of implementing cluster wide procs. The same doc can be found on www.openssi.org website. The paper talks about how to implement cluster wide proccess model without requiring home/deputy concept. But yes it require some core kernel changes. But should be Conditionally enabled like selinux. So overhead for non cluster users should be nill.

I am personally not interested in making intrusive kernel changes even if it yields in true "single-system image". I want very small changes (preferrably none).

Quoted 1 line
Quoted 6 lines
Quoted 4 lines

Regarding my work you can see the status here
http://git.openssi.org/~kvaneesh/gitweb.cgi?p=ci-to-linus.git;a=summary

It only gets the ICS changes. That means it introduce a transport independent kernel cluster framework. Right now it supports two interconnect IPV4 and infiniband verbs.

We also have transport abstraction layer and transport plugins for TCP/IP, SDP (Infiniband and possibly others), and SCI (Dolphin).

I would really like to see this code. Is it available on the web some where ?

Quoted 7 lines
Quoted 5 lines

I am planning on taking the CFS changes. That should bring in clusterwide shared memory too. The way it was done in OpenSSI was to hook a new nopage() function for CFS so that when we page fault, we bring the pages from other node.So i am not sure whether one need a VM hook for getting clusterwide shared memory. But without seeing the code i am clueless.

Nopage will be called if there is no pte. That means, with just nopage you cannot implement RO-RW transition. If you use nopage only, you cannot have multiple readers, because you cannot invalidate all other readers if one reader goes read-write. Thus nopage allows single reader or single writer whle the page fault hook allows multiple readers and single writer.

With the nopage changes CFS also have a token based synchronization mechanism. The token code now works at the granularity of file. But very well can be made to work with file data ranges (range tokens).

To kind of explain how it works with CFS, When you want to map a page you request token server and token server grant
you token. Now somebody when trying to do a write map will result in the above mapped page being unmapped and shipped to the other node. So as long as all are readers there can be multiple nodes sharing the same page.

When one of the node is in writer mode, Then token server will force to unmap the page from other nodes when ever there is a page access.

All these code is opensourced as a part of OpenSSI project. BTW I am only saying that the code at www.openssi.org may be of interest to you.

-aneesh

#10 lore raw
In reply to Constantine Gavrilov

Hi!

Quoted 11 lines

I have implemented SSI (single system image) clustering extensions to Linux kernel in the form of a loadable module.

It roughly mimics OpenMosix model of deputy/remote split (migrated processes leave a stub on the node where they were born and depend on the "home" node for IO).

The implementation shares no code with Mosix/Open Mosix (was written from scratch), is much smaller, and is easily portable to multiple architectures.

Quoted 8 lines

We are considering publication of this code and forming an open source project around it.

I have two questions to the community:

1) Is community interested in using this code? Do users require SSI product in the era when everybody is talking about partitioning of machines and not clustering?

Yes... Remember that some people run hypervisors to enable process migration.

Quoted 4 lines

2) Are kernel maintainers interested in clustering extensions to Linux kernel? Do they see any value in them? (Our code does not require kernel changes, but we are willing to submit it for inclusion if there is interest.)

I'd say so.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html