Vault 2017

linux block multiqueue has been implemented to provide best performance for modern multi-queued IO subsystems like NVMe. The SCSI stack also has been converted to make use of it by providing a simple 1-queue implementation.

However, most existing HBAs like LSI/Broadcom mpt3sas or PMC aacraid already employ MSI-X interrupt steering to spread out the load during interrupt processing, But this is only for the reception side, not the submission side.
So it doesn't quite match with the premises of block-mq, which assumes queue pairs, one for the submission and one for the reception side.

In this presentation I will be giving a performance analysis of the various methods on how block-mq can be mapped onto these HBAs, with the aim of providing tuning hints for the best possible performance and establish the best overall default parameters.