70 lines
2.0 KiB
Plaintext
70 lines
2.0 KiB
Plaintext
Test Author: Pedro
|
|
If it breaks: Pedro
|
|
|
|
Spreader Requote Performance test
|
|
|
|
Note: This test is fairly cumbersome to setup, because performance is
|
|
very sensitive to a number of factors including compilation switches,
|
|
hardware used, thread allocation and load on the machine. Pedro has
|
|
the best handle on how do all this and should probably be the one
|
|
doing this test.
|
|
|
|
|
|
1) Make sure you are testing a release server build
|
|
|
|
2) Make sure you use real hardware, with hyper threading turned off
|
|
and cstates disabled, for the server side
|
|
|
|
3) Make sure you isolate the CPU on which you are running the
|
|
server. This is best done on Linux by editing /etc/default/grub like
|
|
this:
|
|
|
|
# this isolates CPU's 0 to 5 for better performance testing
|
|
GRUB_CMDLINE_LINUX="isolcpus=0-5"
|
|
|
|
4) Now make sure your configuration is set for the lowest latency,
|
|
including dedicating a thread to the order manager on the isolated
|
|
cores:
|
|
|
|
OrdermanCfg {
|
|
value {
|
|
executor {
|
|
thread_name: "oman_orders"
|
|
poll: true
|
|
hosted_actor: "MDSource"
|
|
hosted_actor: "OrderManager"
|
|
}
|
|
executor {
|
|
thread_name: "oman_slow"
|
|
poll: false
|
|
hosted_actor: "*"
|
|
}
|
|
}
|
|
}
|
|
|
|
Threads {
|
|
value {
|
|
thread_group {name_rx: "mdp3feed" core: 0}
|
|
thread_group {name_rx: "oman_orders" core: 1}
|
|
thread_group {name_rx: "jf_ar_sim" core: 2}
|
|
thread_group {name_rx: ".*" core:6 core:7 core:8 core:9 core:10 }
|
|
}
|
|
}
|
|
|
|
5) Now you are ready to start testing. Start a shell and create a new
|
|
synthetic say with 2 CL futures. Use the near month (fast ticking)
|
|
future as the non quoting leg, so that the other leg gets quoted
|
|
often. Make sure to disable throttling in the synthetic definition
|
|
dialog.
|
|
|
|
6) Place an order for this synthetic, away from the market.
|
|
|
|
7) Now run for say 3 minutes. Visit http://youserver:8000/perf_trace
|
|
You should see a graph indicating the median requote times for the
|
|
order you created in 5). Make sure you have at least 100 data points
|
|
and data the median time is around 12us. If you get anything greater
|
|
than 15us, the test has failed.
|
|
|
|
|
|
|