sc-haskell

A survey on the topic of sequentially consistent Haskell.

Inventory of docker images:

First, note that "-dbg" variants count the number of storeLoadBarrier calls and trace them to the ghc-events trace.

Also, "-opt" builds are supposed to have an optimization where we "trusted" the Handle and Bytestring uses of IORef (only).

• v0.2: our first usable dockerized version. ??more here??

• v0.3: (Jun 3) Michael updated the LLVM code generator. ??more here?? Includes the May 20 changes to inline barriers.

• v0.4: (July 12) we did a lot of benchmarking with this one. Several commits. ??

• v0.5: (July _) This was where we fixed the problem with the non-opt version lacking the Handle barriers. In "v0.5" Handle has barriers.

QUESTION: Our earlier numbers from Nofib were 0.3% geomean slowdown (out of line barriers), and 0.1% geomean slowdown for inline barriers.

Now [2016.08.01] we're getting slower nofib results and trying to figure out why.

Notes:

[2016.04.04] {Initial counts for Hackage and stackage}

• 180 bechmark suites among 1770 Stackage packages.

haskell-rebuild-base-hello-world

[2016.05.15] {Benchmarking script}

• Using the benchmarking script in this repo, 134 Stackage LTS-5.16 repos were found to have at least one benchmark. Of those, 66 ran to successful completion

[2016.07.11] {Detective work on bytestring and Sandi}

Suspects:

• Bytestring.append

.