Burstable Cloud Block Storage with Data Processing Units
Talkat 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI '24) · 2024-07-10..07-12·
summary
An end-to-end system called BurstCBS runs on DPUs to balance per-core load with added FPGA queues, uses a two-tier memory pool to prefill buffers without blowing up memory, and schedules I/O with a vectorized I/O cost estimator that models CPU, DMA and FPGA usage separately. The paper claims up to 85% average latency reduction and 5× base-tenant throughput under congestion, pays a 5–8% drop in maximum burst capacity, and validates benefits on MySQL, RocksDB and Alibaba RDS.
AI-written summary · about
chapters (16)
- 0:00Burstable cloud block storage
- 1:27Burstable virtual discs matter
- 2:08Max provisioned IOPS and bandwidth
- 2:53Backend resource availability
- 3:58Interference with burst usage
- 4:43X DPU storage virtualization
- 5:25Congestion on X DPU
- 6:28BurstCBS IO scheduling
- 7:14Mitigating load imbalance
- 7:56Memory for high IOPS
- 8:41Fixed vs dynamic allocation
- 9:26BurstCBS scheduler
- 10:32Vectorized IO cost on X DPU
- 11:13BurstCBS evaluation
- 12:19Performance isolation trade-offs
- 13:02Impact on database systems
related talks
24mModelling Postgres Performance Degradation on Burstable Cloud Instances
48mAnalytics Architectures on Microsoft Fabric
13mBMC: Accelerating Memcached using Safe In-kernel Caching and Pre-stack Processing
25mDisaggregating Persistent Memory and Controlling Them Remotely
17mBreakthrough Data-Centric Computing with a New Memory Tier
43mUse dbt and DuckDB Instead of Databricks for Data Processing
06mAchieve predictable performance
57mNo Downtime, No Limits