[Beowulf] LSF Job Preemption & Checkpointing
raysonlogin at gmail.com
Thu Mar 1 03:22:16 EST 2012
On Thu, Mar 1, 2012 at 2:20 AM, Jan Wender
<j.wender at science-computing.de> wrote:
> AFAIK at least LSF has this as a feature called preemption.
IMO, LSF has the best job preemption & checkpointing support, with the
least integration effort needed from the end user & cluster
administrator. And resource preemption and license preemption are the
more advanced features of LSF.
(There are more manual configuration needed for Grid Engine & Open
Grid Scheduler and/or other batch systems - not impossible, but needs
knowledge on how to tune the scheduler.)
> That depends probably mostly on the application. If the application offers
> it, then the batch system can use it to save state.
> I don't know much about kernel level checkpointing, though.
There are 3 types of checkpointing supported by LSF:
Kernel level is easy, the OS kernel handles everything for the user
(for interactively processes) & the batch system (for jobs).
However, only IRIX, Cray UNICOS, and NEC SUPER-UX support kernel-level
On Linux, you usually need to patch the kernel:
- "Checkpoint/restart: it's complicated": http://lwn.net/Articles/414264/
- "Kernel-based checkpoint and restart": http://lwn.net/Articles/293575/
(Lots of discussions on kernel-level checkpointing in the past few
years but still we don't have anything in the official tree yet...)
Or even kernel assisted user-level checkpointing:
- "Preparing for user-space checkpoint/restore":
And there is also the famous Berkeley Lab Checkpoint/Restart (BLCR),
which is a kernel module and thus you can use your distribution's
- "RCE 12: BLCR": http://www.rce-cast.com/Podcast/rce-12-blcr.html
- "Checkpointing under Linux with Berkeley Lab Checkpoint/Restart":
For user-level, you will need to link against a checkpointing library
shipped with LSF, which (I think) has some object file level init
routines that perform initializations to properly save the state of
stuff and also need to wrap around standard libc functions & system
calls (I forgot the actual details, lots of academic papers published
15 years ago and I recall reading a few of them, but just don't recall
the content :-D ).
See "Standalone Checkpointing":
With user-level checkpointing & restart, you usually need to relink
your application (unless you use the LD_PRELOAD trick). So for
operating systems that don't support kernel-level checkpointing (ie.
most of the OSes), user-level checkpointing usually works for most
general applications (I *think* Platform Computing even ported the LSF
checkpointing library to Windows as well - or at least that's what I
For application-level checkpointing, the applications will handle
everything. But of course each application needs to have its own
built-in support for checkpoint & restart.
Open Grid Scheduler / Grid Engine
Scalable Grid Engine Support Program
> Cheerio, Jan
> ---- Company Information ----
> Vorstand/Board of Management: Dr. Bernd Finkbeiner, Michael Heinrichs,
> Dr. Roland Niemeier, Dr. Arno Steitz, Dr. Ingrid Zech
> Vorsitzender des Aufsichtsrats/Chairman of the Supervisory Board:
> Philippe Miltin
> Sitz/Registered Office: Tuebingen Registergericht/Registration Court:
> Registernummer/Commercial Register No.: HRB 382196
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.
More information about the Beowulf