Is it possible / reasonable to use VMs for the OSS component(s)? I?m wondering about the feasibility of spending my limited storage budget on iSCSI hardware and using some capacity from our existing VMware infrastructure for OSS, and MDx. Something about this seems amiss though. Is it critical to commit ?true? hardware to managing the various devices? This setup would probably not exceed 20Tb spread across 5-6 iSCSI machines (for OSTs). -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.lustre.org/pipermail/lustre-discuss/attachments/20121111/98986909/attachment.html
Hi Jon, It really all depends on the kind of performance expectations you have, you should be able to run VM''s as OSS''s without issue, however the question is why? In most cases you want to utilize various physical nodes to get around single node CPU / back end IO bandwidth limitations. Another possible use case for this is for an active / active standby pair on a single node (running in VM''s) in such a case this may help against localized software failures (logic errors) but won''t do a think to help you with hardware faults. So really you have to consider what you plan to do with this and remember that at some point you''re paying for resources anyways. -cf On 11/11/2012 04:43 PM, Jon Yeargers wrote:> > Is it possible / reasonable to use VMs for the OSS component(s)? I?m > wondering about the feasibility of spending my limited storage budget > on iSCSI hardware and using some capacity from our existing VMware > infrastructure for OSS, and MDx. > > Something about this seems amiss though. Is it critical to commit > ?true? hardware to managing the various devices? This setup would > probably not exceed 20Tb spread across 5-6 iSCSI machines (for OSTs). > > > > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss
I?ve heard that one vendor is basically doing this in their storage appliance, running virtualized Lustre OSS under KVM. We tried tests of this in the FermiCloud project a couple of years ago, running Lustre under KVM and there was a significant performance hit. But since then the KVM I/O system has been improved quite a bit and there have been several new releases of Lustre and the associated server kernels as well, and a new Enterprise Linux major release as well. Probably time to try again. We were interested in doing it because we had a large number of machines with 10TB local storage and 8 cores and we figured that it would not take all 8 cpu cores to serve that much disk and we wanted to make the balance available for compute virtual machines. What did not work at all in those days was to run the Lustre OSS on bare metal and then run other VM?s on top of it. The virtualization was compiled out of the server kernel that was distributed at that time. Steve Timm From: lustre-discuss-bounces at lists.lustre.org [mailto:lustre-discuss-bounces at lists.lustre.org] On Behalf Of Jon Yeargers Sent: Sunday, November 11, 2012 5:43 PM To: lustre-discuss at lists.lustre.org Subject: [Lustre-discuss] Using VMs for OSS machines Is it possible / reasonable to use VMs for the OSS component(s)? I?m wondering about the feasibility of spending my limited storage budget on iSCSI hardware and using some capacity from our existing VMware infrastructure for OSS, and MDx. Something about this seems amiss though. Is it critical to commit ?true? hardware to managing the various devices? This setup would probably not exceed 20Tb spread across 5-6 iSCSI machines (for OSTs). -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.lustre.org/pipermail/lustre-discuss/attachments/20121112/b2c16e52/attachment.html