Increased NVMe performance with BFQ scheduler enabled

Benchmarks Aerospike Server Benchmarks
1

Hi everyone,

Inspired by recently published Phoronix article: https://www.phoronix.com/scan.php?page=article&item=linux-420-io&num=1, I’ve decided to run some benchmarks in our AS lab, I believe that initial results are interesting enough to share:

After setting IO scheduler for SSDPE2KX080T8 NVMe drives (from “none” to “bfq”, with low_latency option enabled), read ops/sec performance reported by asloglatency jumped from ~12200 op/s to ~15500 ops/s (kernel 4.15 no scheduler vs 4.20 BFQ scheduler).

I wonder if anyone on the forum tried it as well and can confirm improved performance?

I can share more details regarding our hardware/benchmark setup if anyone’s interested.

I haven’t run any ACT tests yet to verify my findings.

2

I thought the scheduler would be bypassed if you were giving the raw devices to Aerospike like device /dev/nvme0n1p1 ? Were you using files? This is intriguing :slight_smile: