Permalink
Fetching contributors…
Cannot retrieve contributors at this time
35 lines (27 sloc) 1.4 KB
This plugin supports checkpointing of applications over InfiniBand.
(By default, DMTCP provides checkpoint-restart for TCP (socket-based networks.)
Currently this plugin supports the standard OFED verbs API for InfiniBand,
but vendor-specific InfiniBand APIs are not supported.
This is the work of Jiajun Cao et al. Please report bugs to
dmtcp@ccs.neu.edu as well as to jiajun using his account at ccs.neu.edu.
[
* Note that this plugin will be reorganized by the end of January, 2014.
* The reorganized plugin will not be available in time for the DMTCP 2.1
* release. If you are using DMTCP 2.1, consider getting the newer plugin from
* the DMTCP svn.
]
To use InfiniBand, you must have first configured DMTCP to use it:
To do so:
cd <DMTCP_ROOT> # replace with wherever your DMTCP_ROOT is.
./configure --enable-infiniband-support
make clean
make
bin/dmtcp_launch --ib ./a.out
EXPERTS ONLY: libibtrace.so can also be built
Note, configuring with '--enable-infiniband-support' will still allow
DMTCP to be used with standard TCP instead of InfiniBand.
However, this option is currently incompatible with a DMTCP optimization.
DMTCP normally creates zero-fill-on-demand pages when it finds a
page of all zeroes. The '--enable-infiniband-support' disallows this
optimization. Zero pages will be saved as full pages of zeroes in the
checkpoint image. This is a bug that will be fixed in a future release.