File Access Prediction

Note: "File Access prediction" is part of my research work. There are speed gaps between processors and I/O, network bandwidth restrictions (especially for wireless mobile networks), disconnected network applications (such as mobile computing and AFS), and other reasons, this results in data access latency. Io and the network become performance bottlenecks.

To solve this problem, cache and pre-Fetch are usually used. The cache technology uses the time locality of data access to temporarily retain accessed data. However, due to the size of the cache space and the update algorithm constraints, when data is frequently updated, the cache performance improvement is no longer significant. Prefetch is an active cache technology that uses the spatial locality of data access to predict future data requests. It extracts and caches data before access for user access, this reduces access latency.

Prefetch can improve the system performance, but there are also some shortcomings:
(1) Impact on normal data access load
(2) Incorrect prefetch will reduce the overall system performance
(3) It is difficult to accurately predict and complete prefetch as needed

The basis of prefetch is prediction, which solves the problem of prefetch or something. Why is prediction effective? On the one hand, data requests are not completely random and are driven by user or program behavior. On the other hand, there is an internal association between many data. The key to prediction is to mine data access behavior patterns and build associations between data.

The basic idea of prefetch is to construct an appropriate prediction model by analyzing the data ontology or historical access records, predict the future access mode, and retrieve and cache the data before access, for user access, reducing access latency.

Two Methods of prefetch:
(1) prediction-based prefetch (explicit): precise prediction model, minimizing Error Prediction and smooth prefetch.
(2) grouping-based prefetch (implicit): build a stable Data Group, suitable for relatively stable access modes.

Prediction involves statistical analysis, machine learning, artificial intelligence, and databases.

(Liu aigui/aiguille. Liu)