UADK multi-queue scheduling problem about uadk HOT 2 CLOSED

Do these tests in a test branch (https://github.com/hzhuang1/uadk/tree/debug_sec_v0.1). In this test branch, there's a patch to dump the duty cycles in different periods in each transmission.

commit 582608f0b62e802eb3aab672f0e69bfc432e7de3
Author: Haojian Zhuang <[email protected]>
Date:   Fri Sep 10 14:24:44 2021 +0800

    debug: check the duty cycle
    
    Signed-off-by: Haojian Zhuang <[email protected]>

With this patch, it could dump four periods. They're pick next ctx, lock in algorithm, cipher send, cipher recv.

And there's another patch to remove the mutex in UADK scheduler.

commit e87ce6d78d6fb24a7cb85221c9b0c3038c9e0374 (HEAD -> debug_sec_v0.1, hzhuang1/debug_sec_v0.1)
Author: Haojian Zhuang <[email protected]>
Date:   Sun Sep 12 08:04:38 2021 +0800

    sched: remove the lock
    
    Use lock-free to select one ctx from region.
    
    Atomic operations are used to make sure each operation on region->last
    variable is atomic. Since other global variables are treated as contant
    value. And updating local variables won't change variables in other
    threads.
    
    Signed-off-by: Haojian Zhuang <[email protected]>

hzhuang1@ubuntu:~/uadk$ ./zip_run.sh
start UADK benchmark test.
[--algname]: aes-128-cbc
[--mode]: 1
[--optype]: 0
[--syncmode]:0
[--pktlen]: 1024
[--seconds]: 3
[--thread]: 4
[--multi]: 1
[--ctxnum]: 16
[--algclass]:cipher
[--acctype]: 0
[--engine]:
hisi sec init Kunpeng920!
algname: length: perf: iops: CPU_rate:
aes-128-cbc 1024Bytes 853373.8KB/s 833.3Kops 399.64%
duty cycle:9.08%, 6.46%, 15.58%, 66.27%
start UADK benchmark test.
[--algname]: aes-128-cbc
[--mode]: 1
[--optype]: 0
[--syncmode]:0
[--pktlen]: 1024
[--seconds]: 3
[--thread]: 4
[--multi]: 1
[--ctxnum]: 24
[--algclass]:cipher
[--acctype]: 0
[--engine]:
hisi sec init Kunpeng920!
algname: length: perf: iops: CPU_rate:
aes-128-cbc 1024Bytes 856045.1KB/s 835.9Kops 399.62%
duty cycle:9.43%, 6.16%, 16.18%, 66.99%
start UADK benchmark test.
[--algname]: aes-128-cbc
[--mode]: 1
[--optype]: 0
[--syncmode]:0
[--pktlen]: 1024
[--seconds]: 3
[--thread]: 4
[--multi]: 1
[--ctxnum]: 32
[--algclass]:cipher
[--acctype]: 0
[--engine]:
hisi sec init Kunpeng920!
algname: length: perf: iops: CPU_rate:
aes-128-cbc 1024Bytes 856809.9KB/s 836.6Kops 399.64%
duty cycle:8.59%, 6.76%, 15.51%, 66.39%

With the new data, we could observe all data are similar while ctx numbers are changed. CTX numbers should be always great than thread numbers.

from uadk.

OK

from uadk.