There are two ways to think of SSDs in system design. One is to think of SSD as an extension of disk, where it plays the role of making disks fast and the other is to think of them as an extension of memory, where it plays the role of making memory fat. The latter makes sense when persistence (non-volatility) is unnecessary and data is accessed over the network. Even though memory is thousand times faster than SSD, network connected SSD-backed memory makes sense, if we design the system in a way that network latencies dominate over the SSD latencies by a large factor.
To understand why network connected SSD makes sense, it is important to understand the role distributed memory plays in large-scale web architecture. In recent years, terabyte-scale, distributed, in-memory caches have become a fundamental building block of any web architecture. In-memory indexes, hash tables, key-value stores and caches are increasingly incorporated for scaling throughput and reducing latency of persistent storage systems. However, power consumption, operational complexity and single node DRAM cost make horizontally scaling this architecture challenging. The current cost of DRAM per server increases dramatically beyond approximately 150 GB, and power cost scales similarly as DRAM density increases.
Fatcache extends a volatile, in-memory cache by incorporating SSD-backed storage.
SSD-backed memory presents a viable alternative for applications with large workloads that need to maintain high hit rate for high performance. SSDs have higher capacity per dollar and lower power consumption per byte, without degrading random read latency beyond network latency.
Fatcache achieves performance comparable to an in-memory cache by focusing on two design criteria:
Minimize disk reads on cache hit
Eliminate small, random disk writes
The latter is important due to SSDs' unique write characteristics. Writes and in-place updates to SSDs degrade performance due to an erase-and-rewrite penalty and garbage collection of dead blocks. Fatcache batches small writes to obtain consistent performance and increased disk lifetime.
SSD reads happen at a page-size granularity, usually 4 KB. Single page read access times are approximately 50 to 70 usec and a single commodity SSD can sustain nearly 40K read IOPS at a 4 KB page size. 70 usec read latency dictates that disk latency will overtake typical network latency after a small number of reads. Fatcache reduces disk reads by maintaining an in-memory index for all on-disk data.