Pure Storage is adding Optane read caching to its high-end FlashArray//X systems to make drive-level read accesses up to five times faster.
The //70 and //90 R2 models can be fitted with 750GB Optane 3D XPoint storage-class memory drives, called DirectMemory Modules (DMM). These come in 4- or 8-packs. V5.3.1 of the Purity OS is required.
The DMMs dynamically cache frequently accessed data, which reduces read latency significantly for cache hits. A Pure FAQ says “workloads that make the best use of the cache (100 per cent) hit rate) may see up to a 50 per cent decrease in service latency, which is in addition to SAN and queue latency.” No service or overall latency microsecond numbers are supplied.
Latency can be as low as 250 μs, according to a FlashArray//X datasheet but this does not say if this is service- or overall latency. Blocks & Files understands an Optane SSD has a 10 μs read latency. An NVMe SSD has a 100 to 110 μs read latency.
Pure’s DirectFlash module has about 50 μs latency. The company’s DirectFlash Fabric – end-to-end NVMe – has 200-300 μs latency. The DMM reduces DirectFlash module latency with a read cache hit taking 10 μs – instead of the 50 μs to get the data from DirectFlash, so saving up to 40 μs on read access.
No formal cache sizing method is available yet. The Pure FAQ says a back-of-the-envelope sizer is available for internal use only. The results are not a performance guarantee and are meant only to give Pure employees a pointer to whether a customer would benefit from a cache.
A full sizer will be available in the fourth 2019 quarter as an internal tool, as well as within Pure1 Planning.
There is no cache effect on write latency, bandwidth or any increase in system memory capacity. The DMM cache should be specifically sized for the workload, as adding cache capacity beyond the workload size brings no further latency reduction.
If a single DMM unit fails in a multiple DMM setup, the cache continues to works at reduced capacity.
DMMs should be available by the end of the month by customers whose use case has been qualified by an internal Pure PM team.