Performance & efficiency at Pinterest: Optimizing the latest instances
summary
Pinterest and Intel show concrete results: replacing gzip with zlib-ng and selective zstd use plus gen-over-gen instance upgrades yields up to 50% improvement in cost efficiency and 30–40% per-cluster gains, while a Rockstore fix cuts CPU by ~10%. Benchmarks use perf/PMU profiling, JDK 17 SSL improvements, and instance moves (c5d→c6id, i3→i4i, M7i) and hardware features like AMX and QAT to reach 3.5x encryption throughput and large reductions in CPU usage.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:49Cloud computing trends
- 4:36Microservices and performance
- 8:25Pinterest case study: Merced
- 17:06Pinterest case study: Kafka
- 21:43Pinterest case study: Rockstore
- 23:58Latest generation benchmarks
- 28:19Generative AI and LLMs
- 31:11Metal instances and accelerators
- 35:40Software collaboration and wrap-up
related talks
22mGraviton
33mReimagining the Modern Data Center
36mThe journey of silicon innovation at
42mHow Intel & AWS solve some of the world’s most urgent issues
25mAdopt AWS Graviton2 instances for best price-performance in EC2
01h 01mEstimate your Graviton ROI for improved price performance & sustainability
25mSustainability at the Core with Cloud Native Processors
32mBuild without limits with AWS Cloud infrastructure