Please enable JavaScript.
Coggle requires JavaScript to display documents.
ublk perf - Coggle Diagram
ublk perf
observation
machine
keven Wolf's machine
beaker VM
t/io_uring -p 1
perf is very bad
'fio write/read 64k' is bad
't/io_uring -p 0'
4k bs not good enough
64k bs is good
ublk performs better in 4k IO, but worse in 64K IO
memory bandwidth is saturated with 64KB IO
facts
hardware
laptop vs. server
numa node
frequency
laptop supports higher freq
l1/l2/l3
server is better
communication between block client and ublk_daemon
investigation
check list
cpu utilization
memory bandwidth
bottle neck
numa