NVMe and Multi-MFS support in Ezmeral?

[Michael_Segel]

Hi,

I was going thru the docs and noticed that it had said that MapR doesn't support NVMes. I found this weird because it does.

From past experience ... due to the performance of SSDs and NVMes, to get the most out of MapR-FS, you would want to run multi-mfs and limit 2 SSDs per file system instance.  I was told that for NVMes you'd actually want to split the device and run a file system instance per 1/2 of the device.  (While this may be a bit overkill... I would suspect using one device per file system instance would still give great performance.

Thoughts about using NVMe for a high speed data fabric?

[tdunning]

Putting data on NVMe devices is definitely supported. 

And, as you say, multi-MFS should typically be enabled if you want to get more than a few G Bytes/s out of the system. If a few GB/s sounds fast, you can probably stop reading at the simplest installation. 

The topic of performance with NVMe flash is very device specific, however, and depends on a lot of things like the write intensity that the device was designed for. Involving support or professional services to configure a large-scale high-performance system is definitely recommended.

Can you send me a pointer via email to where you saw the statement that NVMe isn't supported?

[Michael_Segel]

Hi Ted, Thanks for the response. I'll follow up separately. With respect to using Multi-MFS... On my test cluster, the PoCs don't require a lot of data, and the machines can support multiple NVMe drives. I don't know if there's any documentation of MapR-FS and NVMe, but from past experience, I've seen 2 SATA SSD drives per file system instance. With NVMe, I seem to recall splitting the drive. For my servers, I'd rather sacrifice some performance and go w 1 NVMe per instance. That said, is there any documentation in doing this under Epic and installing the MapR containers? Thx

[tdunning]

I have no idea whether multi-mfs would be supported right now with the data fabric running under the container fabric. The container fabric managed instances are more intended as temporary data storage. Longer term storage should have an external instance.

[Michael_Segel]

Right.

This goes back to an earlier discussion I had with you about containerization of the persistent store.

There should be a longer discussion on this,  there are now a lot of options for the persistence layer and creating a true data fabric between compute and persistence.