]> Git Repo - qemu.git/commit - migration.c
rdma: introduce capability x-rdma-pin-all
authorMichael R. Hines <[email protected]>
Wed, 26 Jun 2013 01:35:36 +0000 (21:35 -0400)
committerJuan Quintela <[email protected]>
Thu, 27 Jun 2013 00:38:36 +0000 (02:38 +0200)
commit60d9222c8f50c3e5dd3df9ee84ddd1d1c4b35389
treee15ec28f02b7d995171274aa47703244880dd827
parent43487c678d6e4e7182bfa70d2bc75422578782aa
rdma: introduce capability x-rdma-pin-all

This capability allows you to disable dynamic chunk registration
for better throughput on high-performance links.

For example, using an 8GB RAM virtual machine with all 8GB of memory in
active use and the VM itself is completely idle using a 40 gbps infiniband link:

1. x-rdma-pin-all disabled total time: approximately 7.5 seconds @ 9.5 Gbps
2. x-rdma-pin-all enabled total time: approximately 4 seconds @ 26 Gbps

These numbers would of course scale up to whatever size virtual machine
you have to migrate using RDMA.

Enabling this feature does *not* have any measurable affect on
migration *downtime*. This is because, without this feature, all of the
memory will have already been registered already in advance during
the bulk round and does not need to be re-registered during the successive
iteration rounds.

Reviewed-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>
Reviewed-by: Chegu Vinod <[email protected]>
Reviewed-by: Eric Blake <[email protected]>
Tested-by: Chegu Vinod <[email protected]>
Tested-by: Michael R. Hines <[email protected]>
Signed-off-by: Michael R. Hines <[email protected]>
Signed-off-by: Juan Quintela <[email protected]>
include/migration/migration.h
migration.c
qapi-schema.json
This page took 0.026586 seconds and 4 git commands to generate.