This message was deleted.
# _general
s
This message was deleted.
d
We're dealing with that with more RAM soon enough. I'm just concerned that the newer and supposedly better hosts are performing dramatically differently.
k
I’m just saying I’ve seen this behavior before. Also, without exact configs you can’t tell. If pure is using 400 spindles and HF 2.. yeah, you’ll get better perf from pure.
t
Disk latency can be tied to many factors, including the RAID configuration, network tuning (if over iSCSI or NFS), lack of server resources, etc. etc. It's often quite the quest to find the bottleneck. Queue depth can be exceeded if too many VMs are assigned to a single logical storage unit. And on and on....
d
Oh- ditto, I've seen this for years and that's why I harp on cache sizing. My concern is more that the exact same VMs configured the same way with the same workloads balanced between them are so much worse. I'm asking specifically around Hyperflex since this is not something I've dealt with a lot.
t
Overhead, you'd think, would not be that different. Typically, iowait latency comes from poor RAID performance, which can be in either reads, writes, or both. If writes, it's usually because the RAID isn't efficient enough to process the write requests. The iostat utility in Linux is useful for monotoring meterics like that and if the queue is saturated, that will also produce the same bottleneck.
k
I’m not a HF expert but I’d be very surprised if they’re relying on HW RAID given the HCI nature
t
It'll affect the peformance of the array whether HW or SW RAID regardless, if the RAID configuration sucks.
k
Technically you’re correct but some HCI solutions aren’t using HW/SW raid as they’ll get their data availability by the software defined storage layer. Could be the same for HF..
t
Absolutely, Kees. SDS generally does a great job for reads, not necessarily for writes, hence my being curious which side (reads or writes) or both are affected,
n
I have an all flash local storage vsan and when turning on mcsio during catalog creation the vm's all had much disk latency as well, not sure why though. before mcsio latency is 1-2, afterwards was upwards of 20ms for the exact same workload so i ditched it.
d
Well- MCSIO helps with writes a little bit, but if you don't give it enough RAM it doesn't work well at all. Either way, you're going to have a lot of reads with MCS. PVS shouldn't have much of anything going on tho; and it doesn't. I'm just wondering about the latency
k
Careful with that statement on ‘SDS’ and writes, most modern day HCI Solutions can deal with write intensive data streams (some have SAP HANA very, thats 9GB/s data throughput when the DB flushes, that’s no joke)