Unsolved
This post is more than 5 years old
21 Posts
0
1062
July 5th, 2011 11:00
Binding FAST Cache mirrors between DPE and DAE disks
I am configuring two VNX5300 arrays where each array will have four EFDs for FAST Cache. My initial thought was to split the two RAID1 mirrors so each mirror has a disk in the DPE -- so in the very remote chance that a DAE is lost due to a major hardware failure I would avoid whatever "badness" would happen if a FAST Cache mirror is lost with dirty write cache data. All of my other pools and RGs will be able to sustain a DAE loss (RAID10 with mirrors across DAE's and buses); obviously if the DPE is lost I'm having a bad day/week.
According to page 72 in the "EMC Unified Storage Best Practices for Performance and Availability Common Platform and Block Storage 31.0" document linked in another thread, you should not split RAID 1 groups including FAST Cache drives across the DPE and a DAE. The underlying reasoning is hinted at avoiding a rebuild on boot in case of total power failure.
Is the loss of an entire DAE so remote (dual PSU failure, etc) that I should just lose the tinfoil hat? Is the only drawback to splitting the mirror between DPE/DAE going to be the subsequent IOPS and utilization hit for a potential rebuild once the array is powered back on after a power failure?
*edit - Going by the best practice guide, the design would then have one of three config choices: all four EFD's in the DPE; one mirror in the DPE and the 2nd mirror in a DAE; or all four EFDs in two separate DAEs. The last two would be to balance the EFD mirrors between buses. Both arrays will have at least three DAE6S' enclosures.
*edit2 - I guess I could also split the mirrors between DAEs if I don't put any EFDs in the DPE. This would balance out the bandwidth between buses as well as prevent the rebuild issue from a complete power loss. The only question would be how bad do things get if FAST Cache EFDs are not in the DPE and power is lost?
0 events found
jps00
2 Intern
•
392 Posts
0
July 5th, 2011 12:00
>Is the loss of an entire DAE so remote (dual PSU failure, etc) that I should just lose the tinfoil hat?
Lose the hat.
The loss of a DAE with all its built-in redundancy would leave a larger problem than a slow boot. (Read the section on how the DAEs are daisy-chained together.
> Is the only drawback to splitting the mirror between DPE/DAE going to be the subsequent IOPS and utilization hit for a potential rebuild once the array is powered back on after a power failure?
Its a question of scale. Your two Flash drives in a VNX5300 rebuilding out of the DPE are not going to saturate the SAS back-end port. If your storage system had five or more drives rebuilding, there would be a different discussion.
> the design would then have one of three config choices:
Your reading of the Best Practices is correct about the choices. However, to make it simpler, with flash drive-based RAID groups and the DPE/DAE O/S you are, "either all in, or all out".
HTH.
Noogz
21 Posts
0
July 5th, 2011 13:00
It looks like I can go "all out" with the FAST Cache disks by putting them in two DAEs and also keep my tinfoil hat on by splitting the mirror pairs between the DAEs. How 'bout them apples?
Given the odds it may not do anything for me, but at least I'll feel better at night.
dynamox
9 Legend
•
20.4K Posts
0
July 5th, 2011 13:00
how many buses are in 5300 ? If there are four, would it make sense to put each EFD on a bus of its own ?
dynamox
9 Legend
•
20.4K Posts
0
July 5th, 2011 14:00
cool ..i have an eval 5700, thought i saw more than 2.
Noogz
21 Posts
0
July 5th, 2011 14:00
The 5300 only has 2 buses. Only the 5700 and above have more -- the 5700 has 4 and 7500 has 4 or 8.
jps00
2 Intern
•
392 Posts
0
July 6th, 2011 04:00
> It looks like I can go "all out"
That is the way I'd do it.
Sometimes you have to make compromises when deciding between performance and availability. They compete for the same resources. Generally I lean toward performance, because failures are the exception, and not the rule.