Back to index

Broadcast Disks: Data management for asymmetric communication environments.

S. Acharya, R. Alonso, M. Franklin, S. Zdonik. This summary also includes material from Dissemination based data delivery using broadcast disks, same authors minus Franklin.

One-line summary: Multilevel (but not truly "hierarchical") carouseling of multiple broadcast "disks" onto a single channel; constructing the program and managing the client caches must be considered in tandem, since not all uncached pages are equidistant from client. A caching strategy of LRU-per-disk works well in practice.

Overview/Main Points

Relevance

Flaws

I think more attention should be focused on application level considerations: what is a "page" from the point of view of typical apps? How do large "logical pages" get subdivided into physical pages (application level framing) and can you exploit application semantics to determine how to schedule the fragments? Can you analyze application usage/traffic to help construct the bcast schedule? I recognize that this paper represents early work so maybe the authors will do this as they relax their simplifing assumptions.

Also the relationship to NUMA work was mentioned in passing but I'm surprised they weren't able to draw more from it, or at least state the differences (e.g. duality between compiler-directed static data placement for NUMA programs and constructing a static broadcast schedule).


Back to index