iopsystems/rezolus v5.27.0

on GitHub

Changed

• tcp_packet_latency costs less per segment and per read. It kept
  each socket's start time in a hash map: every data segment attempted an
  insert, taking the bucket lock, and a read that found an entry deleted it.
  It now keeps them in a fixed 4 MiB array of one-word slots indexed by a
  hash of the socket's address, and attaches tp_btf programs where the
  kernel has bpf_skc_to_tcp_sock (5.9+), raw_tp otherwise. In a 6.12 VM
  over loopback (valkey-benchmark 4 KiB SET/GET, 48 clients, one run per
  arm), ns per run for tcp_probe / tcp_rcv_space_adjust went from
  393 / 598 to 254 / 364 with tp_btf and to 296 / 307 with raw_tp.
  • Only an in-order data segment (sequence number rcv_nxt) starts a
    sample. Duplicate, out-of-order and beyond-window segments used to, and
    a dropped one could leave a start time that a much later read recorded.
    An in-order segment the kernel then drops still starts one, as before.
  • A segment whose slot holds another socket's start time is not timed:
    about K / 2^19 of segments when K sockets hold a start time at once
    (0.20-0.22% of samples at 2000 clients against an exact reference).
    When a later in-order segment arrives before the read, the sample
    starts there and is shorter than the data's wait.
  • The old map stopped timing new sockets once 10240 held a start time;
    misses now grow with K instead (about 2% at K = 10240).