AuriStor File System Ubik Services at Scale
Presenter:
- Jeffrey Altman, AuriStor, Inc.
Mike Kazaar’s 1989 IEEE paper “Ubik: Replicated Servers Made Easy” describes the performance goals of Ubik services as supporting one write transaction per second and ten to one hundred read transactions per second on multiple servers (for load sharing and availability purposes.) AuriStor File System cells simultaneously service many tens of thousands of cache managers accessing hundreds of thousands of volumes maintaining many hundreds of thousands of rx connections and hundreds of thousands of user/process and host identities.
Marcio Barbosa’s 2019 talk “Overcoming Ubik Limitations” described some of the technical hurdles to obtaining reliable Ubik performance. This talk will describe the AuriStorFS Ubik implementation that has sustained 40,000 read transactions per second across a quorum of six servers while averaging 22 write transactions per second continuously for more than two days.
The AuriStorFS Ubik service achieves these performance numbers while prioritizing data integrity and correctness. Ubik application services were originally designed around the light-weight process (lwp) model. Safely scaling a Ubik application service requires not only altering the Ubik protocol and library implementation but also rethinking application level caching and data access rules.