Replacement strategies for XQuery caching systems

Abstract

To improve the query performance over XML documents in a distributed environment, we develop a semantic caching system named ACE-XQ for XQuery queries. ACE-XQ applies innovative query containment and rewriting techniques to answer user queries using cached queries. We also design a fine-grained replacement strategy which records user access statistics at a finer granularity than the complete XML query regions. As a result, less frequently used XML view fragments are replaced to maintain a better utilization of the cache space. Extensive experimental results illustrate the performance improvement achieved by this strategy over the traditional one for a variety of situations.

Introduction

Due to the growing demand by web applications for retrieving information from multiple remote XML sources, it has become increasingly critical to improve the efficiency of XML query evaluation. One key step towards achieving such an optimization is to exploit caching technology to reduce the response latency caused by data transmission over the Internet. Inspired by the semantic caching idea [14], which utilizes cached queries and their results to answer subsequent queries by reasoning about their containment relationships, we propose to build such a caching system to facilitate XML query processing in the Web environment.

One major difference between semantic caching systems [6], [14], [21] and the traditional tuple [16] or page-based [6] caching systems is that the data cached at the client side of the former is logically organized by queries instead of physical tuple identifications or page numbers. To achieve effective cache management, the access and management of the cached data in a semantic caching system is thus typically at the level of query descriptions. For example, the decision of whether the answers of a new query can be retrieved from the local cache is based on the query containment analysis of the new query and the cached query descriptors themselves, rather than by looking up each and every tuple or page identification of objects that could possibly answer a current user request.

The semantic caching idea has been extensively studied in the relational context [14]. However, query evaluation and containment dealing with XML data differ in their nature and difficulty from those in the relational setting. New challenges are being imposed by the tree-oriented nature of XML and the XQuery language on the tasks of query containment and rewriting, as we will point out in this paper.

We have developed the first XQuery-based caching system, named ACE-XQ [9], [11], to deploy our proposed query containment and cache management techniques in the XML context. In ACE-XQ, new and cached queries are both expressed in XQuery, a quickly thriving XML query language proposed by W3C as the standard [44]. The query descriptors in the ACE-XQ system help to capture the query semantics which are utilized in the decision for query containment. While [9] describes the XQuery containment and rewriting techniques in ACE-XQ, in this paper we focus on cache management of ACE-XQ, in particular cache replacement issues.

Typically, a cache system utilizes a replacement manager to decide what to retain in the cache and what to discard in case of a full cache. In a query-based caching system, the data granularity for replacement is the query and its associated query result. The cache manager in ACE-XQ maintains a collection of query regions, each composed of a query descriptor and the corresponding XML view document, i.e., query region=query descriptor + result XML view. Query descriptors can be utilized for reasoning about the containment relationships between the cached queries and the new query. Also, user access statistics information may be attached to the query descriptors by the deployed replacement strategy to calculate the region utility values. The replacement manager usually picks the cached query with the lowest utility value and purges it to make room for the new query.

Since a new query is often conceptually subsumed by or overlapping with previously cached queries, the query region of the latter can be seen logically segmented into two pieces. One corresponds to the overlapping part which is to be retrieved by the probe query for answering the new query. The left-over piece does not contribute to answering the new query. The replacement manager of a traditional query-based caching system may split the containing query region into two regions corresponding to their respective usefulness in this latest query answering process. After the splitting, a uniform utility value is then maintained for each query region. Whenever the cache is full, a complete query region would be the unit for replacement. However, such a region-splitting scheme entails a large decomposition overhead each time when a new query overlaps with the cached queries. Also, it would result in more and more smaller XML view documents over time which are possibly less useful in answering future queries due to their fragmentation.

An alternative solution is to tolerate some redundancy in the cached queries. That is, even if newly incoming queries partially overlap with existing queries, we would opt to not split existing queries in order to avoid fragmentation. Then a straightforward application of replacement would be to replace a complete query region at each iteration. However, the data granularity of a whole query region being deleted each time in such a replacement strategy may be too coarse for “large” XML views. This would impact the cache space utilization. Also, such a replacement strategy does not reflect the contribution of different fragments in a cached XML view which may participate in answering different subsequent queries. Replacement at the granularity of complete XML views hence suffers apparent drawbacks.

We now propose a refined replacement strategy, namely, to record utility values for finer regions of existing cached views in terms of their internal structure rather than assigning a uniform value for the whole cached query region [12]. To be precise, we attach to each query descriptor a detailed path table listing all paths returned in the query. When a cached query contains or partially overlaps with a new query, the utility statistics of those paths requested by the probe query are updated, however without splitting the cached query. When the cache is full, the replacement manager does not select complete regions but only specific paths with the lowest utility value within such query regions for replacement. It then composes a filter query to remove the fragments corresponding to those paths from the cached XML view. The relevant query descriptors are then modified accordingly to be consistent with the changed XML view.

This proposed partial replacement strategy utilizes the view structure to maintain utility values at a finer granularity than complete query regions. This way, the replacement helps to maintain in the cache the most likely “hot” query regions. This is because the original cached queries may be refined by future filter queries that remove the less useful fragments within them. It hence forgoes the explicit region splitting upon every new incoming query, avoiding the generation of too many small region fragments with little use for answering future queries.

We have also implemented both the proposed partial replacement strategy as well as the complete region replacement strategy (which we now call total replacement) within our ACE-XQ caching system. In this paper, we now report upon the extensive experimental study we have conducted to compare the performance of our partial replacement and the alternative total replacement strategies in a variety of scenarios. The results show that in most cases especially when the cache size is medium, the partial replacement strategy outperforms the total replacement strategy in terms of hit count ratio (HCR), hit byte ratio (HBR) and query response delay.

The rest of the paper is organized as follows. In Section 2, we show the running example queries to motivate the need for a query containment and rewriting solution in the context of XQuery. An overview of our overall XQuery caching solution ACE-XQ is given in Section 3. We then focus on the cache management aspect of the ACE-XQ system. We analyze the advantages and disadvantages of alternative query region managing schemes in Section 4, while in Section 5 we describe a fine-grained replacement strategy (a la partial replacement) deployed in ACE-XQ. The experimental studies comparing our partial replacement strategy with the traditional total replacement strategy are given in Section 7. The related work is described in Section 8 and we conclude in Section 9.