[LU-16640] parallellize generic I/O

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Assignee: Patrick Farrell (Inactive)
• Reporter: Patrick Farrell (Inactive)

Description

LU-8964 tried to improve I/O performance by splitting I/Os to multiple threads at the ll_file/cl_io level (ie, the top level).  This is a decent idea, but had some problems.

Specifically, the original LU-8964 parallel I/O work didn't improve performance much and in some cases harmed it (mostly because of limitations in the buffered I/O path, and the synchronous nature of the DIO path at the time), the new parallel DIO path should make it possible to get large performance improvements with roughly the same approach as in LU-8964.  It's mostly a matter of using this or not intelligently - We will only do parallel IO on the DIO path, since it's mostly a net loss on the buffered path.  We will also add some sized based tuning.

But the patch will be modeled on the approach taken in LU-8964.

Issue Links

• is related to: LU-16540 move ll_get_user_pages to ptlrpcd threads