Unified Buffer Management (UBM)
Unified Buffer Management classifies file-block accesses and allocates cache space among sequential, looping and other-reference partitions.
Detection, replacement and allocation are separate modules. Per-file consecutive blocks identify sequences; repeated sequences establish loops. Sequential blocks use MRU, loops use period-based replacement with MRU ties, and the paper uses LRU for other references. [1]
For a loop of l blocks with period p, the model gives Hit(n)=min(l,n)/p. Its marginal gain is 1/p through n=l and zero beyond saturation. Time here counts all block references, not seconds. Comparing partitions requires the same observation clock. [1]
In a constructed four-block loop with period 20, two buffers yield 2/20=0.1 hits per global reference, although 2/4=0.5 of that loop's accesses hit in the model. Doubling the period to 40 halves the former rate without changing the latter fraction. A fifth buffer adds no hits after all four loop blocks fit. [2]
A newly detected sequence initially lacks loop evidence. Under cache pressure, its first repeated access can miss before the detector recognizes looping; cold-cache spare space can conceal this cost. Detection therefore does not promise a first-reuse hit. [1]
Relations
Limitations
- Estimated reference-count gains do not price unequal I/O latency, dirty writeback or prefetch costs. [1]
Citations
- Kim et al. A Low-Overhead High-Performance Unified Buffer Management Scheme that Exploits Sequential and Looping References. OSDI 2000, sections 3-5.
- Hesela. Loop-allocation arithmetic and explicit global-reference denominators.