Low latency programming is increasingly important across a variety of use cases. Still, many of the tips and tricks of low latency are only part of developer folklore. This document attempts to codify that knowledge for people to (re)discover the art of low-latency programming.
- Latency is a distribution
- Avoid coordinated omission
- Co-locate compute and data e.g. Processing-In-Memory or Processing-Near-Memory
- Replicate data for faster access
- Maximize cache hit rate
- Control memory access patterns
- Avoid dynamic memory management
- Avoid demand paging to prevent memory thrashing e.g. by using larger memory pages (hugepages on Linux, superpages on FreeBSD, ...)
- Avoid as much work as possible (for example, avoid function call overhead by using inlining)
- Avoid CPU intensive computation.
- Partition data to avoid sharing (and, therefore, synchronization)
- Make shared data structures read-only (when possible)
- Reduce head-of-line blocking
- Avoid context switching
- Use wait-free data synchronization
- Use busy-polling instead of wakeups
- Disable Nagle's algorithm
- Use non-blocking I/O
- Parallelize requests to different services
- Request hedging (send redundant requests to multiple replicase, use response from fastest one)
- Use optimized SIMD instructions for suitable problems
- Multiprocessing and multithreading
- Use preemptible kernel
- Interrupt and process affinity
- Watch out for bad device drivers
- Use kernel-bypass networking such as DPDK or XDP
- Use hardware offload with accelerators and FPGA
- 11 Best Practices for Low Latency Systems by Ben Darfler (2014).
- Optimizing web servers for high throughput and low latency by Alexey Ivanov (2017).
- The Tail at Scale by Jeffrey Dean and Luiz André Barroso (2013)
- Tales of the Tail: Hardware, OS, and Application-level Sources of Tail Latency by Jialin Li et al (2014)
- Amdahl’s Law for Tail Latency by Christina Delimitrou and Christos Kozyrakis (2018)