Filesystems and Distributed Caches
Keeps a globally distributed client-side cache strongly consistent by handing out and revoking cache tokens.
- Filesystems
- Distributed caching
- Consistency
Work done at HPE Labs, Systems Architecture Lab. Caching on the client is easy until you need every client to agree on what it is caching. CacheFS is a filesystem that keeps a globally distributed client-sided cache strongly consistent. Published at HPE TechCon ‘26 (82 of 1036 submissions accepted). Two related inventions are currently under internal disclosure review, with patent filing on approval.
What it sits on
Building the filesystem meant first understanding what the layers beneath it actually cost:
- FUSE. Linux’s Filesystem in Userspace turns out to lose meaningful performance to kernel scheduling decisions. Characterising those slowdowns produced a set of optimisations now under development, and a prototype RDMA-based in-memory filesystem on top of FUSE used to test them.
- eBPF. Instrumentation is not free, and the standard practice of measuring a system with eBPF probes perturbs the thing being measured. Probe Less, Learn More characterises that overhead and corrects for it. To appear at the ProTools workshop at SC ‘26.
- Message latency. Characterising sustained low-level message latency, presented at the Cray User Group ‘26.
There is a lot that can be added here. In due time :)