Redesigning Storage Systems for Future Workloads Hardware and Performance Requirements

https://www.youtube.com/watch?v=UFxS2fepBLk

Redesigning Storage Systems

• New: applications, hardware, performance requirements

• Applications

  • Billions of impressions collected by most web services for analytics

  • Past workloads: read mostly, rarely updated

  • Future workloads: example of IoTs

    • Mixed workload

    • Read / write ratio

    • Reads: read large volumes of sensor, data for control and analytics

    • Writes: new sensor data, continuously recorded

  • We produce huge amount of data: grow exponentially

  • Need to figure out a way to serve data in persistent storage

  • Nature of the data is quite diverse: unstructured data

    • KVs: key to big, unstructured data

    • API: put(), get(), and scan()

    • Example: RocksDB, redis, cassandra

• Hardware

  • Landscape: RAM, HDD, Tape (archive)

  • Now: NVM, SSD between layer of RAM and HDD; and Glass, DNA, ...

  • Focus: SSD

    • New SSDs are much faster

    • Random accesses almost as fast as sequential accesses

• Performance requirements

  • Throughput

    • reads & writes / second

    • higher is better

    • should be steady

  • Latency

    • time it takes to read or write

    • lower is better

    • tail latency (e.g., 99th percentile)

      • High tail latency in KVs --> unpredictable performance. Difficult to provide QoS guarantees

Contributions

• Applications: cluster metadata management. Mixed write-intensive workload.

  • TRIAD [ATC '17]: decrease maintenance I/O to increase client throughput.

    • HyperLogLog-driven compaction

    • WAL repurposing

    • Hot/Cold key separation

• Hardware: new servers with ample memory sizes (100s GB)

  • FloDB: Scale KVs with memory size

    • New 2-level data structure

    • Mostly O(1) insert

    • Concurrent reads and updates

• Performance requirements

  • Need low, steady tail latency. No latency spikes.

  • SILK: KVs I/O scheduling

    • Opportunistic I/O bandwidth allocation for KV maintenance ops

    • Prioritize client work

    • No tail latency spikes

• KVell: fast KV for NVMe SSDs (SOSP 19')

  • RocksDB: low, unstable throughput. CPU-bound on Fast SSDs.

    • Can't saturate the I/O bandwidth that comes with the device

  • Existing KVs internals