Quoting in your message looks kind of messy so forgive me if I’m propagating that below.
Honestly I agree that the Optanes will give diminishing returns at best for all but the
most extreme workloads (which will probably want to use NVMoF natively anyway).
>>
>> This does split up the NVMe disk into 4 OSDs, and creates WAL/DB
> partition on the Optane drive - however, it creates 4 x 223 GB partitions
> on the Optane (whereas I want 35GB partitions).
Because you want to use the rest of the space on the Optane for something else?
>>
>> It talks there about "osd op num shards" and "osd op num threads
per
> shard" - is there some way to set those, to achieve similar performance to
> say, 4 x OSDs per NVMe drive, but with only 1 x NVMe? Has anybody done any
> testing/benchmarking on this they can share?
I’d like to see that too, since I’m curious if this is not still limited by per-OSD
serialization. Those options have been around for years haven’t they?
I’m also curious how splitting an NVMe device into multipe OSDs could affect write
amplification.
— aad