October 2012 Summaries

We investigated how fork times influence performance on different AWS platforms and Redis dataset sizes. Fork is an expensive operation in Unix-like systems, causing delays in other operations. We tested Redis 2.4.17 on various platforms, including Xen-based instances, to find that fork time per GB of memory improves with stronger instances, but still causes significant latency. In contrast, the Redis Cloud minimizes fork times through several mechanisms and technologies, providing high performance at affordable infrastructure prices, even when dataset sizes grow.

To address the challenge of running Redis with persistence storage on AWS, a recent study was conducted to assess the impact of persistent storage on performance. The study found that, despite concerns about slow EBS and AOF being more demanding than snapshotting, running Redis with persistence storage using AOF with "fsync every second" policy does not significantly affect Redis' performance under normal conditions. In fact, the average response time was only 1-8% higher with data-persistence on most platforms, and the difference was negligible when using a non-raided EBS configuration. The study also found that the optimal EBS configuration of the Redis Cloud platform results in better performance, especially for smaller instances where shared servers can lead to significant latency gaps. Overall, the study suggests that AOF with "fsync every second" policy should be considered as the default Redis configuration for users running on AWS, and the Redis Cloud's architecture provides a unique advantage in terms of performance and cost-effectiveness.

We compared the performance of Amazon's PIOPS EBS with Standard EBS for Redis applications running on AWS. We tested three configurations: PIOPS EBS with Optimized-Instance, Standard EBS with non-Optimized Instance, and Standard EBS with non-Optimized Instance. Our results showed that neither configuration outperformed the other in terms of RPS (Requests Per Seconds) and Latency. However, we found that using a specific standard EBS configuration can provide equal or better performance than PIOPS EBS, especially when it comes to sequential writes. We also calculated the cost implications of each configuration and found that Standard EBS is generally cheaper. These findings are based on a specific storage usage pattern, so results may vary depending on different storage access patterns.