JinSXS wrote:
if the recommendation is to put SAS+NVME disks in the same CPG, won't there be a inconsistent
performance/latency skew due to the different interface ?
notice there is no NVME SCM also, so we can't do like AFC of NVME SCM with SAS SSD right
im very surprise that primera isn't a full NVME capable on the base/controller chassis
As mentioned in previous post, I don't think there is a different in throughput/iops performance between the drives. As for latency you get the latency of the slowest drives.. so NVMe only CPG probably may provide lower throughput and lower latency. I think most people will take 100% higher throughput vs 5-10% lower latency (already well below 1msec). Remember that best practice is "what works best for most", not the answer to everything.
Remember that AFC was read only. You can get a Primera 670 with 1TB cache per node which will also do write. Sounds like a better option
![Smile :)](./images/smilies/icon_e_smile.gif)
.
Full NVMe in base chassis is probably a simple math question. Each NVMe slot uses 4 PCIe lanes. 4x24 = 96 PCIe lanes. Not sure what CPU and PCIe lanes available is in the Primera, but I'm guessing that after dedicating PCIe lanes to ASICs, HBAs (FC/iSCSI/SAS), RCIP ports etc that there is nowhere near 96 left..... my guess is that if they made a solution for 24 NVMe slots in node chassis, they also solved how to get NVMe to expansion cages which they haven't yet.
But I'm guessing this is the evolution. First get NVMe as a proven technology in production systems. Then make it scale before SCM (or similar) flash technology becomes mainstream and the real difference between NVMe and SAS is visible.