Re: [ovirt-devel] NUMA-support oVirt scheduler intergation design
Hi,
After some thinking, since memory weights aren't supported yet in the scheduler,
I will exclude weights for now (will add a RFE for NUMA weights that will depend on Memory
weights RFE).
Also, to avoid several memory filters, will extend current memory filter to support pinned
virtual (guest) NUMA nodes
(filter out hosts that cannot accommodate memory of VM's pinned virtual NUMA nodes
within a host's physical NUMA nodes).
Please review the patch: http://gerrit.ovirt.org/#/c/28467/
Thanks,
Gilad.
----- Original Message -----
From: "Gilad Chaplik" <gchaplik(a)redhat.com>
To: "Doron Fediuck" <dfediuck(a)redhat.com>, devel(a)ovirt.org
Cc: "Shang-Chun Liang (David Liang, HPservers-Core-OE-PSC)"
<shangchun.liang(a)hp.com>, "Chegu Vinod"
<chegu_vinod(a)hp.com>, "Xiao-Lei Shi (Bruce, HP Servers-PSC-CQ)"
<xiao-lei.shi(a)hp.com>, "Chuan Liao (Jason Liao,
HPservers-Core-OE-PSC)" <chuan.liao(a)hp.com>
Sent: Wednesday, June 4, 2014 1:23:48 PM
Subject: NUMA-support oVirt scheduler intergation design
Hi List,
This is the design for NUMA-support oVirt Scheduler integration; I'll add 3
policy unit modules (#) as described in following document:
(1) Filter Module: NUMA_HostFilter:
If the VM has NUMA vNodes/pinning configured, and the rule can't be
accommodated on the host, it'd filtered.
NOTES:
* pin2host filter should make sure the filter gets a single host. the
filter should be transparent to specific hosts.
* not sure we'll want to add this filter in first phase to all
cluster/scheduling policies; @Doron?
Weights Modules:
(2) even_dest_NUMA
Should "even distribute" weights (scores) according to:
* Available memory on nodes
* Numatune memory mode (strict/preferred/interleave)
* Autonuma/pinned NUMA
(3) power_saving_NUMA (added to power_saving cluster/scheduling policy)
* same as even_dest_NUMA with opposite weights.
I'd appreciate any comment or suggestion.
Once approved by you, I'll update the feature page and upload a patch for
review.
Thanks,
Gilad.