Accelerate the Image Distribution for DADI Using Intel DSA
Talkat ContainerCon @ Open Source Summit Europe 2022 · on Friday, 2022-09-16· at 16:50· Wicklow Hall 2A (Level 2), Convention Centre Dublin, Spencer Dock, N Wall Quay, North Wall Dublin 1, D01 T1·
summary
overlaybd presents remote container images as a virtual block device and does checksum (CRC) on-stream, where CRC costs about 10% of work; moving CRC into the NIC/accelerator reduces CPU and memory pressure. using Intel DSA to compute CRC during DMA drops bad blocks before they touch RAM, offloads checksum cycles from cores, and integrates with SPR platform features. implementation uses overlaybd with LZ4 chunks sized to sector granularity, replaces SSE4.2 CRC with DSA ops, and shows a measured 9% overall performance increase versus the previous Alibaba baseline.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:03Study Tabi
- 2:07Open Source
- 3:04White Block
- 6:42Darkboxes
- 7:42Collaboration with Intel
- 9:44Implementation
- 11:18Intel DSA
- 16:42Example
- 17:14Software Architecture
- 18:17Performance Tests
- 19:23Project Status
- 20:12Feature Work
- 21:02Future Work
- 21:45Thank You
related talks
17mDSA-2LM: A CPU-Free Tiered Memory Architecture with Intel DSA
34mThree Approaches to Speed up Image Distribution in Cloud Native Era
16mWhen DDIO Meets Page Coloring: Revisiting DDIO Performance with Sepia
29mDC-Store: Eliminating Noisy Neighbor Containers using Deterministic I/O Performance
27mSpeeding Up Linux Disk Encryption
42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
17mPara-ksm: Parallelized Memory Deduplication with Data Streaming Accelerator
35mIntroducing UACCE Bus of DPDK