You may remember a post from over a year ago titled Analysis of SSHFS performance for large builds, in which I outlined how Google exposes its gigantic source monorepo via a FUSE file system and in which I analyzed the performance of large builds using SSHFS to access such file system.
As part of those experiments, I played with pCacheFS, a Python-based FUSE file system that provides a persistent caching layer on to top of a slow mount point. Pure SSHFS benchmarks were poor performance-wise, and pCacheFS benchmarks were not better. In the former case, the issues were because of the lack of caching and single-threaded server operation, and in the latter because of slow performance of a CPU-bound multi-threaded Python app.
As a result, I wrote a new file system to act as a persistent caching layer on top of an SSHFS mount, and I called that file system sourcachefs (note the missing e for a cool pun). sourcachefs implements the basic same idea behind pCacheFS but in Go for efficiency, with more features, and with a lot of performance tuning into it.
Further experiments with sourcachefs did show up performance improvements in the builds involved, but sourcachefs was never deployed widely within Google for complex reasons I won’t get into (plus we got a much better solution to the problem we were originally facing). At this point, sourcachefs is just a personal project of mine that has no involvement with my daily job, but very soon you’ll see the publication of a different project that shares a lot in common with sourcachefs!
I digress. Today, after over a year of finding little chunks of time during long flights to clean up the code base, it has become minimally worthy of publication… so I have pushed the code to GitHub. Would be a pity for it to be lost, even if it’s just my first Go incomplete experiment.
Head to http://github.com/jmmv/sourcachefs/ for details.
— Julio, from DL439 at 36000 feet.
