Low crypto performance of AES-NI MB PMD on multi-core

[Koji Ikeura]

Hi,

I am using dpdk-test-crypto-perf to measure the crypto performance of AES-NI MB PMD.
The performance reports say "The SW PMD performance is linear scaling out with core numbers",
but there are no result on multiple cores and this is not the case in my environment.
What about your environment? Does anyone know how to improve it?

With DPDK v22.07/intel-ipsec-mb v1.20 and DPDK v22.11/intel-ipsec-mb v1.00 the performance
did not scale out with the number of cores, but with DPDK v17.05.2/intel-ipsec-mb v0.45
the performance do scaled out with the number of cores.

Results and environment are below.

# Results (DPDK v22.07 / intel-ipsec-mb v1.20)
- 8 cores
$ ./dpdk-test-crypto-perf -l 0-8 --vdev crypto_aesni_mb_pmd -- --ptest throughput --devtype crypto_aesni_mb --optype cipher-then-auth --cipher-algo aes-cbc --cipher-op encrypt --cipher-key-sz 16 --auth-algo sha1-hmac --auth-op generate --auth-key-sz 64 --digest-sz 12 --total-ops 10000000 --burst-sz 32 --buffer-sz 64

     lcore id    Buf Size  Burst Size    Enqueued    Dequeued  Failed Enq  Failed Deq        MOps        Gbps  Cycles/Buf

            8          64          32    10000000    10000000           0           0      1.8980      0.9718     1053.74
            7          64          32    10000000    10000000           0           0      1.9231      0.9846     1040.01
            5          64          32    10000000    10000000           0           0      1.8613      0.9530     1074.49
            3          64          32    10000000    10000000           0           0      1.8205      0.9321     1098.62
            1          64          32    10000000    10000000           0           0      1.7988      0.9210     1111.86
            2          64          32    10000000    10000000           0           0      1.7988      0.9210     1111.88
            6          64          32    10000000    10000000           0           0      1.6651      0.8525     1201.14
            4          64          32    10000000    10000000           0           0      1.6618      0.8508     1203.52

- 1 core
$ ./dpdk-test-crypto-perf -l 0-1 --vdev crypto_aesni_mb_pmd -- --ptest throughput --devtype crypto_aesni_mb --optype cipher-then-auth --cipher-algo aes-cbc --ciph
er-op encrypt --cipher-key-sz 16 --auth-algo sha1-hmac --auth-op generate --auth-key-sz 64 --digest-sz 12 --total-ops 10000000 --burst-sz 32 --buffer-sz 64

     lcore id    Buf Size  Burst Size    Enqueued    Dequeued  Failed Enq  Failed Deq        MOps        Gbps  Cycles/Buf

            1          64          32    10000000    10000000           0           0      5.7756      2.9571      346.28

# Environments
- CPU
Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz

- OS
Oracle Linux Server 8.4

- Kernel
4.18.0-240.el8.x86_64

- DPDK and intel-ipsec-mb version
DPDK     intel-ipsec-mb
v22.07   v1.20 <-- do not scale out
v20.11.3 v1.00 <-- do not scale out
v17.05.2 v0.45 <-- scale out

-- 
Koji Ikeura