Optimizing Webservers for High Throughput and Low Latency | Dropbox

paper: Making Linux TCP Fast
- fair queueing and pacing
  - decreases retransmit rates
- congestion control
  - tcp_cubic → tcp_bbr
- ack processing & loss detection
  - use newer kernels
    - new heuristics enabled by default
    - upgrade your kernel
  - sysctl
    - don’t
      - enable time wait recycle
      - disable timestamps unless you know what you’re doing
    - do
      - disable slow start after idle
      - enable TCP mtu probing
      - increase read and write memory - match to BDP
libraries
- lots of time spent in compression?
  - zlib forks by cloud flare & Intel; zlib-gz
  - Malloc?
    - jemalloc
    - tcmalloc
  - pcre symbols in perftop?
    - consider JIT
- TLS
  - openssl, libressl, boringssl
  - symmetric encryption
    - big files?
      - openssl symbols in perftop
        
        enable chacha20-poly1305
nginx
- what to compress
- when to compress
  - brotli for static content as a build step
  - optimize for round trip time
    - encrypt, transmit, decrypt
- buffering
  - x-accel-buffering
- TLS
  - session resumption
    - sticky load balancing or distributed cache
    - tickets - no server side state
  - TLS record sizes
- event loop stalling
  - aio_write
global optimizations
- load balancing
- inbound & outbound traffic engineering
- Twitter: @SaveTheRbtz

bagrounds.org