Alas, in spite of nvme website saying that the specs are released, https://nvmexpress.org/specifications/ hasnt been updated in >18 months. Looking forward to getting to read in, once these become available.
[Ed: Oh! Subsystem Local Memory Commands and Computational Programs both shows up since I first checked.]
> Enabling live migration of PCIe NVMe controllers between NVM subsystems.
A little confusing to me, but they want to be able to suspend a vm with nvme and Live Migrate that virtual drive somewhere else. Not sure what the problems with this today were but ok?
> New host-directed data placement for SSDs that simplifies ecosystem integration and is backwards compatible with previous NVMe specifications.
This sounds potentially related to Zoned Namespaces? That has been the basket that everyone put their eggs in after the "let the host do it all" Open Channel Flash folks kind of gave up, it feels like.
Hypothetically this is supposed to be much less work, just writing into blocks of storage & deleting the while block. But so far there have been only one or two drives with this capability, afaik, and you can't just go buy them.
Haven't seen any videos or resources on this host directed topic. The wins we could get by doing less are so huge; hoping there's some overlap with these Open Channel flash ideas.
- Support for offloading some host processing to NVMe storage devices.
Very interesting to see where this starts! On-drive data processing can get a real start now. The nvme video talks about, for example, doing CRC processing.
https://youtu.be/w3eiI7IUM8k#t=16m51s
More advanced work (dedicated hardware accelerators, not software) & not NVMe (CXL), but look at Marvell's just announced CXL memory controller Structera, which has hardware lz4 encoder/decoders.
The more basic use case is likely to be query pushdown, putting one filtering directly on the ssd. As noted in the nvme video above, it's actually not that fast in singular (since hard drives don't have 3GHz manycores chips on them) but with lots of drives it can aggregate to big wins.
NFS can already use RDMA. The real killer app here is to be able to load data from storage directly into GPU memory and avoid polluting CPU caches and funneling this data through the PCIe root complex straw.
[Ed: Oh! Subsystem Local Memory Commands and Computational Programs both shows up since I first checked.]
> Enabling live migration of PCIe NVMe controllers between NVM subsystems.
A little confusing to me, but they want to be able to suspend a vm with nvme and Live Migrate that virtual drive somewhere else. Not sure what the problems with this today were but ok?
> New host-directed data placement for SSDs that simplifies ecosystem integration and is backwards compatible with previous NVMe specifications.
This sounds potentially related to Zoned Namespaces? That has been the basket that everyone put their eggs in after the "let the host do it all" Open Channel Flash folks kind of gave up, it feels like.
Hypothetically this is supposed to be much less work, just writing into blocks of storage & deleting the while block. But so far there have been only one or two drives with this capability, afaik, and you can't just go buy them.
Haven't seen any videos or resources on this host directed topic. The wins we could get by doing less are so huge; hoping there's some overlap with these Open Channel flash ideas.
- Support for offloading some host processing to NVMe storage devices.
Very interesting to see where this starts! On-drive data processing can get a real start now. The nvme video talks about, for example, doing CRC processing. https://youtu.be/w3eiI7IUM8k#t=16m51s
More advanced work (dedicated hardware accelerators, not software) & not NVMe (CXL), but look at Marvell's just announced CXL memory controller Structera, which has hardware lz4 encoder/decoders.
The more basic use case is likely to be query pushdown, putting one filtering directly on the ssd. As noted in the nvme video above, it's actually not that fast in singular (since hard drives don't have 3GHz manycores chips on them) but with lots of drives it can aggregate to big wins.