On the Impact of I/O Access Patterns on SSD Storage
Résumé
Nowadays, solid-state drives (SSDs) are increasingly used in storage infrastructures thanks to their promising performance compared with hard disk drives (HDDs). Their internal architecture consists of parallel storage units that do not use mechanical movements, so considerably fortified for random data access. However, in this report, we show that some SSD storage is sensitive to I/O patterns, achieving low I/O performance in case of running random workloads. The performance degradation occurs with workloads that have specific characteristics such as having only one I/O process and I/O requests with small block sizes. After many investigations on the I/O stack and inside the SSD storage, our results show that the internal readahead of SSDs is behind the performance gap between the sequential and random workloads. Indeed, activating the internal readahead incurs a performance increase in case of sequential access thanks to pre-reading the next amount of data into the cache. In contrast, this makes no sense for random workloads as the upcoming requests may not depend on what already stored in the cache.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |