About: Cache is a(n) research topic. Over the lifetime, 59167 publication(s) have been published within this topic receiving 976633 citation(s).
Papers published on a yearly basis
15 Feb 2016
Abstract: Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9x to 13x; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.
••21 Mar 1999
TL;DR: This paper investigates the page request distribution seen by Web proxy caches using traces from a variety of sources and considers a simple model where the Web accesses are independent and the reference probability of the documents follows a Zipf-like distribution, suggesting that the various observed properties of hit-ratios and temporal locality are indeed inherent to Web accesse observed by proxies.
Abstract: This paper addresses two unresolved issues about Web caching. The first issue is whether Web requests from a fixed user community are distributed according to Zipf's (1929) law. The second issue relates to a number of studies on the characteristics of Web proxy traces, which have shown that the hit-ratios and temporal locality of the traces exhibit certain asymptotic properties that are uniform across the different sets of the traces. In particular, the question is whether these properties are inherent to Web accesses or whether they are simply an artifact of the traces. An answer to these unresolved issues will facilitate both Web cache resource planning and cache hierarchy design. We show that the answers to the two questions are related. We first investigate the page request distribution seen by Web proxy caches using traces from a variety of sources. We find that the distribution does not follow Zipf's law precisely, but instead follows a Zipf-like distribution with the exponent varying from trace to trace. Furthermore, we find that there is only (i) a weak correlation between the access frequency of a Web page and its size and (ii) a weak correlation between access frequency and its rate of change. We then consider a simple model where the Web accesses are independent and the reference probability of the documents follows a Zipf-like distribution. We find that the model yields asymptotic behaviour that are consistent with the experimental observations, suggesting that the various observed properties of hit-ratios and temporal locality are indeed inherent to Web accesses observed by proxies. Finally, we revisit Web cache replacement algorithms and show that the algorithm that is suggested by this simple model performs best on real trace data. The results indicate that while page requests do indeed reveal short-term correlations and other structures, a simple model for an independent request stream following a Zipf-like distribution is sufficient to capture certain asymptotic properties observed at Web proxies.
TL;DR: This paper demonstrates the benefits of cache sharing, measures the overhead of the existing protocols, and proposes a new protocol called "summary cache", which reduces the number of intercache protocol messages, reduces the bandwidth consumption, and eliminates 30% to 95% of the protocol CPU overhead, all while maintaining almost the same cache hit ratios as ICP.
Abstract: The sharing of caches among Web proxies is an important technique to reduce Web traffic and alleviate network bottlenecks. Nevertheless it is not widely deployed due to the overhead of existing protocols. In this paper we demonstrate the benefits of cache sharing, measure the overhead of the existing protocols, and propose a new protocol called "summary cache". In this new protocol, each proxy keeps a summary of the cache directory of each participating proxy, and checks these summaries for potential hits before sending any queries. Two factors contribute to our protocol's low overhead: the summaries are updated only periodically, and the directory representations are very economical, as low as 8 bits per entry. Using trace-driven simulations and a prototype implementation, we show that, compared to existing protocols such as the Internet cache protocol (ICP), summary cache reduces the number of intercache protocol messages by a factor of 25 to 60, reduces the bandwidth consumption by over 50%, eliminates 30% to 95% of the protocol CPU overhead, all while maintaining almost the same cache hit ratios as ICP. Hence summary cache scales to a large number of proxies. (This paper is a revision of Fan et al. 1998; we add more data and analysis in this version.).
TL;DR: This paper proposes a novel coded caching scheme that exploits both local and global caching gains, leading to a multiplicative improvement in the peak rate compared with previously known schemes, and argues that the performance of the proposed scheme is within a constant factor of the information-theoretic optimum for all values of the problem parameters.
Abstract: Caching is a technique to reduce peak traffic rates by prefetching popular content into memories at the end users. Conventionally, these memories are used to deliver requested content in part from a locally cached copy rather than through the network. The gain offered by this approach, which we term local caching gain, depends on the local cache size (i.e., the memory available at each individual user). In this paper, we introduce and exploit a second, global, caching gain not utilized by conventional caching schemes. This gain depends on the aggregate global cache size (i.e., the cumulative memory available at all users), even though there is no cooperation among the users. To evaluate and isolate these two gains, we introduce an information-theoretic formulation of the caching problem focusing on its basic structure. For this setting, we propose a novel coded caching scheme that exploits both local and global caching gains, leading to a multiplicative improvement in the peak rate compared with previously known schemes. In particular, the improvement can be on the order of the number of users in the network. In addition, we argue that the performance of the proposed scheme is within a constant factor of the information-theoretic optimum for all values of the problem parameters.
TL;DR: Specific aspects of cache memories investigated include: the cache fetch algorithm (demand versus prefetch), the placement and replacement algorithms, line size, store-through versus copy-back updating of main memory, cold-start versus warm-start miss ratios, mulhcache consistency, the effect of input /output through the cache, the behavior of split data/instruction caches, and cache size.
Abstract: design issues. Specific aspects of cache memories tha t are investigated include: the cache fetch algorithm (demand versus prefetch), the placement and replacement algorithms, line size, store-through versus copy-back updating of main memory, cold-start versus warm-start miss ratios, mulhcache consistency, the effect of input /output through the cache, the behavior of split data/instruction caches, and cache size. Our discussion includes other aspects of memory system architecture, including translation lookaside buffers. Throughout the paper, we use as examples the implementation of the cache in the Amdahl 470V/6 and 470V/7, the IBM 3081, 3033, and 370/168, and the DEC VAX 11/780. An extensive bibliography is provided.
Trending Questions (10)