Is this a reply to Paul’s message from 11 months ago?
https://bit.ly/32oZGlR
The PM1725b is interesting in that it has explicitly configurable durability vs capacity,
which may be even more effective than user-level short-stroking / underprovisioning.
Hi. How do you say 883DCT is faster than 970 EVO?
I saw the specifications and 970 EVO has higher IOPS than 883DCT!
Can you please tell why 970 EVO act lower than 883DCT?
The thread above explains that. Basically it’s not as simple as “faster”. IOPS describe
behavior along one axis under a certain workload for a certain length of times. Subtle
factors:
* With increasing block size, queue depth, operation rate / duration, some less-robust
drives will exhibit cliffing where their performance falls off dramatically
——————
|
|________
—————————————————
(that may or may not render usefully, your UMA may vary)
Or they may lose your data when there’s a power event.
* Is IOPS what you’re really concerned with? As your OSD nodes are increasingly saturated
by parallel requests (or if you’re overly aggressive with your PG ratio) , you may see
more IOPS / throughput, at the risk of latencying going down the drain. This may be
reasonably acceptable for RGW bucket data, but maybe not indexes and for sure not for RBD
volumes.
* The nature of the workload can dramatically affect performance
** block size
** queue depth
** r/w mix
** sync
** phoon
** etc
This is one thing that (hopefully) distinguishes “enterprise” drives from “consumer”
drives. There’s one “enterprise” drive (now EOL) that turned out to develop UREs and
dramatically increased latency when presented with an actual enterprise Ceph — vs desktop
— workload. I fought that for a year and found that older drives actually fared better
than newer, though the vendor denyed an engineering or process change. Consider the total
cost of saving a few bucks on cheap drives that appear *on paper* to have attractive
marketing specs, vs the nightmares you will face and the other things you won’t have time
to work on if you’re consumed with pandemic drive failures.
Look up the performance firmware update history of the various 840/860 EVO even when used
on desktops, which is not to say that the 970 does or doesn’t exhibit the same or similar
issues. Consider if you want to risk your corporate/production data, applications, and
users on desktop-engineered drives.
In the end, you really need to buy or borrow eval drives and measure how they perform
under both benchmarks and real workloads. And Ceph mon / OSD service is *not* the same as
any FIO or other benchmark tool load.
https://github.com/louwrentius/fio-plot
is a delightfully visual tool that shows the IOPS / BW / latency tradeoffs
Ideally one would compare FIO benchmarks across drives and also provision multiple models
on a given system, slap OSDs on them, throw your real workload at them, and after at least
a month gather drive/OSD iops/latency/bw metrics for each and compare them. I’m not aware
of a simple tool to manage this process, though I’d love one.
ymmocv
— aad
_______________________________________________
ceph-users mailing list -- ceph-users(a)ceph.io
To unsubscribe send an email to ceph-users-leave(a)ceph.io