Support handling TCP I/O in simple separate io_uring-based I/O threads

Required mainly for clients, allows to scale parallel client I/O with TCP
from 100-150k iops to ~400k iops and from 2-3 GB/s to at least 7-8 GB/s
with 4 I/O threads, at the same time increasing Q=1 latency by 2x thread
switching delay, which is ~10 us when CPU powersaving is disabled and may
be as high as 200 us when it's enabled.
This commit is contained in:
Vitaliy Filippov
2024-07-04 13:29:20 +03:00
parent 21d1171ba4
commit abbba6ade4
14 changed files with 331 additions and 19 deletions
+18
View File
@@ -1,3 +1,21 @@
- name: osd_iothread_count
type: int
default: 0
info: |
TCP network I/O thread count for OSD. When non-zero, a single OSD process
may handle more TCP I/O, but at a cost of increased latency because thread
switching overhead occurs. RDMA isn't affected by this option.
Because of latency, instead of enabling OSD I/O threads it's recommended to
just create multiple OSDs per disk, or use RDMA.
info_ru: |
Число отдельных потоков для обработки ввода-вывода через TCP-сеть на
стороне OSD. Включение опции позволяет каждому отдельному OSD передавать
по сети больше данных, но ухудшает задержку из-за накладных расходов
переключения потоков. На работу RDMA опция не влияет.
Из-за задержек вместо включения потоков ввода-вывода OSD рекомендуется
просто создавать по несколько OSD на каждом диске, или использовать RDMA.
- name: etcd_report_interval
type: sec
default: 5