Synthetic Performance Testing
We can't fit more than a handful of comparison products in the charts and still keep them tidy and readable. We have to be selective in the process. For this review, we chose the Micron 9100 Max because its DNA leads back to Memblaze. The company also sold a similar model two generations ago. We also chose the Intel DC P4510 8TB, a very popular enterprise SSD used in many data centers today.
Sequential Read Performance
The C916's interface advantage shows best with sequential data at high queue depths. With our server, we achieved roughly 6,300 MB/s with peak performance coming at 32 outstanding IO (OIO). The 2.5" D916 bumped up against the interface's 3,500 MB/s ceiling at 16 OIO while reading 128KB sequential data.
Sequential Write Performance
In steady state, both PBlaze5 SSDs show similar sequential write performance. The only real variation we saw in this workload was at 4 OIO where the add-in card was able to use its bandwidth advantage to surpass 3,500 MB/s
Sequential Mixed Workload Performance
NVMe is a bidirectional interface, and that allows the PBlaze5 D916 to handle mixed sequential data at speeds beyond 3,500 MB/s.
The C916 shows us where it's extra interface performance becomes useful in this test. The C916 is capable of delivering the same sequential performance as the D916 from 40% writes to 100% writes. The C916 outperforms the D916 in the 100% reads to 70% reads areas.
Random Read Performance
The two PBlaze5 SSDs deliver nearly identical random read performance in our test system at all OIO. Both drives trail the Intel DC P4510 slightly until 64 OIO where the performance equals. At higher queue depths the Memblaze SSDs overtake the Intel.
Random Write Performance
The Micron 9100 Max is an older Memblaze design built for Micron. The new PBlaze5 has similar peak performance as the 9100 Max, but at lower OIO, the new design shows a significant advantage.
Random Mixed Workload Performance
The PBlaze5 916 SSDs show just a modest increase over the Micron 9100 Max at 100% workloads (all reads and writes), but the real advantage comes in the middle with mixed random workloads. We rarely see mixed workloads portrayed on specification sheets, but in most cases, this is where your data lives.