2 Replies to “Postgres performance speedups, possibly”

  1. Would be helpful to describe each of the “updates” in the plot, and what they are.

  2. Petr, the various patches were tentative and now these results are dated, we perform even better than this in master. I plan to do a comparative graph of past dragonfly versions as well as perhaps freebsd or linux around release time.

    The big takeaway from this is, green is where we were a few weeks ago, purple is where we were a few days ago, and we are doing even better today.

    The big changes between green and purple are: pool locks were added to kqueue subsystem, making the locking there finer grained and the file descriptor table was made finer-grained also.

    The big changes between then and now are that the netisr’s can be messaged fully asynchronously and we have a new/updated userland memory allocator, dmalloc, which hasn’t hit the tree yet.

Comments are closed.