I received some questions today around the storage latency requirements for vSphere Metro Storage Cluster (vMSC) solutions. In the past the support limits were strict:
- 5ms RTT for vMotion for Enterprise license and lower, 10ms RTT for vMotion for Enterprise plus
- 5ms RTT for storage replication
RTT stands for Round Trip Time by the way. Recently, and today I noticed I never blogged about this, the support limits have changed. For instance EMC VPLEX supports up to 10MS RTT for vMotion (not fully tested for stretched cluster / vSphere HA). Which indeed makes a lot of sense to have it aligned with the vMotion limits as more than likely the same connection between sites is used for both storage replication and vMotion traffic.
So I would recommend anyone who is considering implementing a vMSC environment (or architecting one) to contact your storage vendor about their support limits when it comes to storage latency.
That’s useful… all too often, when people are talking about cold sites and geographical distribution, backups, all at the same time as wanting not a second of data loss, or more than a few seconds of failover time in case of a disaster: I find lots of people seem to totally ignore (or forget) about issues such as latency and concerns such as split-brain (total break in datacenter connectivity between a two site stretched cluster, with none holding quorum is also disasterous). A stretched cluster “sounds so appealing” on the surface, because the theoretical benefits are so great; If it weren’t for ugly little engineering limitations, including the speed of light, and necessary failover delays, every enterprise should have one, otherwise: if the clustering concept itself weren’t a significant risk of failure and data loss.
>5ms may be technically supported, but should it be recommended? Nope.
Especially for virtualized applications, where the added up IO queue latency and synchronous write latency increase due to commit-on-replicate could be significant.
I think folks rapidly forget some risks involved, and it becomes difficult to convince folks that 1, or 10 gigabits of bandwidth across the continent won’t work so well. For now I would say current stretched cluster options seem to be a cute hack at best. [A cool hack, but a hack.]
Then there’s the matter of Layer 2 networking extension resulting in inefficient routing, and extension of certain failure domains across sites in stretched cluster designs — and horrendous new-failure-mode-inducing complications to routing to attempt to correct.
People also think the network must be broke, if they can’t FTP or Rsync a file over TCP between datacenters at 30ms+ latency, and get a transfer at the full 1000 Megabit speed rate of that private link (without realizing the need to adjust protocol choices or buy expensive TCP optimization solution).
And then there’s a lot of the thinking that disk drive performance is MB/sec — latency gets forgotten there too.
“Wow, this must be the best 2TB SATA disk drive ever… I can get 500 MB/sec out of it..
This will be perfect for virtualizing SQL servers… I’ll just put a few of them in RAID5, to improve performance and make sure the data is very safe.”
“At what latency, though?”
“Huh? Latency? I’m sure it doesn’t matter. I can copy a 5 Gig file to it in 10 seconds.”
“How many mean random 4K write IOPS, and what standard deviation?”
“What’s a write IOP?”
See… someone could probably write a book on things that the vendor should have been asked about, but you erroneously assumed weren’t important; or didn’t understand to ask. I would favor proactive vendors — a responsible vendor should inform _you_ of this fact, before they let you buy their product for a stretched cluster