This post is more than 5 years old
58 Posts
0
1181
Caching for directors in same engine
I'm aware of the idea of optimizing cache vs availability depending on how many directors a given host is attached to. The more directors, the more availability, but the increased chance of duplicate data being cached in multiple directors.
For directors in the same engine, whether just a small VPLEX, or one engine in a medium or large configuration, are there any cache optimizations? As in, do directors in the same engine have the ability to "see" or access each others' cache (say, via CMI bus) in a way that directors in different engines do not?
Essentially just trying to understand if there is a difference in caching behavior when looking at Engine1 DirA and Engine1 DirB vs Engine1 DirA and Engine2 DirB. Or if a director is a director, no matter where it is.
garyo
89 Posts
0
March 9th, 2015 08:00
Hi Sean,
On VPLEX CMI was actually disabled as of GeoSynchrony 5.1. There was a defect whereby interrupts for new messages between directors on the same engine would sometimes get lost. As a result, intra-director communication over CMI would sometimes take up to 1 second or longer.
So as of GeoSynchrony 5.1 and newer, all inter-director communication within a cluster is done via the local FC network. And when we disabled the CMI and relied only upon FC for local COM traffic, there was actually very little performance differences.
There's no performance advantages to creating a storage-view with directors on the same engine.
There never was any kind of RDMA between directors in the same engine.
Thanks,
Gary
seancummins
226 Posts
0
March 9th, 2015 06:00
Hi raid-zero,
Yes, if two directors in the same VPLEX engine need to exchange cache information, they'll do so over the CMI instead of going over the COM (internal FC networks).
Incidentally, VMAX3 does the same thing.
Thanks,
- Sean
rzero
58 Posts
0
March 9th, 2015 07:00
Thanks for the quick answer Sean. As a follow up, if this read over CMI happens, does the standard behavior of reading into local dir cache before returning to host still happen? Or does it bypass local cache?
seancummins
226 Posts
1
March 9th, 2015 07:00
Sure thing. Besides transferring over the CMI instead of the COM, I don't believe there is any difference between how we handle a global read hit when the hit is from the peer director in the same engine vs. a director in a different engine -- data is still copied to the local director's cache before being sent to the host.
Thanks,
- Sean
seancummins
226 Posts
0
March 9th, 2015 08:00
Thanks Gary!
Sorry about the outdated info, raid-zero.
Thanks,
- Sean