SIGMETRICS '02 Proceedings of the 2002 ACM SIGMETRICS international conference on Measurement and modeling of computer systems
Workload Modeling for Performance Evaluation
Performance Evaluation of Complex Systems: Techniques and Tools, Performance 2002, Tutorial Lectures
The workload on parallel supercomputers: modeling the characteristics of rigid jobs
Journal of Parallel and Distributed Computing
ARC: A Self-Tuning, Low Overhead Replacement Cache
FAST '03 Proceedings of the 2nd USENIX Conference on File and Storage Technologies
Easy and Efficient Disk I/O Workload Characterization in VMware ESX Server
IISWC '07 Proceedings of the 2007 IEEE 10th International Symposium on Workload Characterization
Buzztraq: predicting geographical access patterns of social cascades using social networks
Proceedings of the Second ACM EuroSys Workshop on Social Network Systems
Finding a needle in Haystack: facebook's photo storage
OSDI'10 Proceedings of the 9th USENIX conference on Operating systems design and implementation
Proceedings of the 20th international conference on World wide web
TailGate: handling long-tail content with a little help from friends
Proceedings of the 21st international conference on World Wide Web
Workload analysis of a large-scale key-value store
Proceedings of the 12th ACM SIGMETRICS/PERFORMANCE joint international conference on Measurement and Modeling of Computer Systems
nsdi'13 Proceedings of the 10th USENIX conference on Networked Systems Design and Implementation
Hi-index | 0.00 |
In a social network, it is natural to have hot objects such as a celebrity's Facebook page. Duplicating hot object data in each cluster provides quick cache access and avoids stressing a single server's network or CPU resources. But duplicating cold data in each cache cluster consumes significant RAM. A more storage efficient way is to separate hot data from cold data and duplicate only hot data in each cache cluster within a data center. The cold data, or the long tail data, which is accessed much less frequently, has only one copy at a regional cache cluster. In this paper, a new sampling technique to capture all accesses to the same sampled keys is created. We then calculate the working set size for each key family for estimating the memory footprint. We introduce an important metric, duplication factor, as the ratio between the sum of each individual cluster's working set size and the regional working set size. We analyze why some key families have a higher duplication factor. It is important to separate hot keys and cold keys from the same key family with minimal overhead. We present a novel cache promotion algorithm based on key access probability. We also proposed a probability model based on the binomial distribution to predict the promotion probability with various promotion thresholds. Our experiment shows by shrinking the cluster level cache layer and having a fat regional level cache for cold data, we are able to achieve a higher combined cache hit ratio.