Corsair Force MP600 PCIe 4.0 NVMe SSD Benchmarks On Linux

[coder]

Phoronix: Corsair Force MP600 PCIe 4.0 NVMe SSD Benchmarks On Linux

It's almost as if this benchmark was rigged to show the drive in a good light. The Flexible I/O Tester & FS-Mark use 2 MB and 1 MB sizes, which are mostly throughput-limited. Why no random 4 kB tests to show its raw IOPS potential?

I can't say much about the DB benchmarks, since it's not obvious how big the records are or how much coherency exists in the access patterns.

[coder]

Yeah that's a known performance deficit for windows but it doesn't explain why the nvme on pci-e 4.0 would suck more on win compared to linux... that's what's troubling me.

Try comparing similar benchmarks on both systems. I'm guessing the Windows benchmarks you saw were much more skewed towards random performance and IOPS-intensive workloads.

[coder]

I'm sorry to say that I wouldn't trust much those results.
tbh I don't trust them at all since the results are pretty much incoherent.

The Samsung 970 pro shouldn't be slower in any test compared to the Samsung 860.

It's just a guess, but I wonder if the benchmark is doing a fsync between files. In that case, the bottleneck could be the NAND, more than the interface.

You're right that it's weird result, and worth understanding, but doesn't necessarily invalidate the tests.

[coder]

I wonder if Optane is bottlenecked by PCIe3 and would see a similar perf improvement using PCIe4

No, the 900P has a PCIe 3.0 x4 interface, which is nominally capable of up to 4 GB/sec. The 900P is bottlenecked by its controller - not the interface. In the 905P, Intel replaced the controller with one of their own design.

[Michael]

It's almost as if this benchmark was rigged to show the drive in a good light. The Flexible I/O Tester & FS-Mark use 2 MB and 1 MB sizes, which are mostly throughput-limited. Why no random 4 kB tests to show its raw IOPS potential?

I can't say much about the DB benchmarks, since it's not obvious how big the records are or how much coherency exists in the access patterns.

There's lots of 4K and more here: https://openbenchmarking.org/result/...HV-CORSAIRFO44

As for the DB tests, all the tests are open and transparent via OpenBenchmarking.org.

[coder]

There's lots of 4K and more here: https://openbenchmarking.org/result/...HV-CORSAIRFO44

Where are the rest of the drives in your article? Those benchmarks don't do me a lot of good, in isolation.