PROCESSOR-SDK-AM62X: In AM62x SDK 10.01.04, the performance of the hardware acceleration engine is lower than that of the software.

zhangxiaohui

Currently, we have upgraded the SDK to version 10.01.04, but testing the hardware acceleration engine on this SDK still reveals issues.

Test Conclusions: AM62X SA2UL Hardware Acceleration Engine

For AES Algorithm:
- The CPU usage is reduced by 50% with hardware acceleration, but the encryption speed is 5 to 7 times slower compared to software implementation.
For SHA Algorithm:
- The CPU usage between software and hardware acceleration is almost the same (99% vs 98%), but the encryption speed with hardware is 2 times slower than software.

Test Command

time -v openssl speed -elapsed -evp aes-128-cbc

Test Data

测试场景	不同字节块的吞吐量（KB/s)						CPU占用率
测试场景	16	64	256	1024	8192	16384	CPU占用率
软件	86,389.33	269,102.23	551,346.18	783,013.55	891,265.02	898,542.25	99%
硬件加速	354.07	1,534.19	5,749.33	22,647.81	90,608.98	117,391.36	44%

AES-192-CBC

测试场景	不同字节块的吞吐量（KB/s)						CPU占用率
测试场景	16	64	256	1024	8192	16384	CPU占用率
软件	83,026.60	242,356.57	457,208.32	607,948.12	672,030.72	676,599.13	99%
硬件加速	355.65	1,539.75	5,749.85	22,657.02	86,466.56	109,996.71	43%

AES-256-CBC

测试场景	不同字节块的吞吐量（KB/s)						CPU占用率
测试场景	16	64	256	1024	8192	16384	CPU占用率
软件	79,876.45	223,361.79	398,315.69	509,919.91	554,713.09	557,733.21	99%
硬件加速	355.03	1,427.95	5,714.35	21,408.09	81,892.69	103,912.79	42%

sha256

测试场景	不同字节块的吞吐量（KB/s)						CPU占用率
测试场景	16	64	256	1024	8192	16384	CPU占用率
软件	11,142.87	41,235.24	133,912.15	303,965.87	485,111.13	505,670.31	99%
硬件加速	422.05	1,682.30	6,543.87	24,064.34	144,924.67	217,858.05	98%

sha512

测试场景	不同字节块的吞吐量（KB/s)						CPU占用率
测试场景	16	64	256	1024	8192	16384	CPU占用率
软件	6,579.98	26,080.51	56,940.54	95,400.62	118,972.42	120,973.99	99%
硬件加速	406.62	1,625.77	5,729.02	19,031.72	51,574.10	58,545.49	98%

The above performance test data is similar to that of the hardware-accelerated OpenSSL in processor-sdk-linux of SDK 10_01_10_04, both significantly lower than the software implementation. Please help determine whether the hardware acceleration engine can be further optimized.

software-dl.ti.com/.../Linux_Performance_Guide.html

4 个月前

0 Alice 4 个月前

TI__Mastermind 35890 points

您好，
已经收到了您的案例，调查需要些时间，感谢您的耐心等待。

0 Daniel 4 个月前

TI__Mastermind 36245 points

您好

Could you please clarify what is attached data again. I see two entries.

0 zhangxiaohui 4 个月前

Prodigy 30 points

The test results are presented in two columns, with the software results in the upper column and the hardware acceleration engine results in the lower one.

0 Daniel 4 个月前回复 zhangxiaohui

TI__Mastermind 36245 points

您好

This is the response from the devlopment team:

Crypto accelerator are actually off loaders and will only save CPU cycles.

They are efficient in cryptographic operations running at 400Mhz but not faster than CPU running at 1.4GHz.

We mostly look at at CPU utilization while checking if offload is happening properly

In case of CPU it is 99%, if offload is happening it is somewhere around 30-66%.

0 zhangxiaohui 4 个月前

Prodigy 30 points

Dear Team,

Thank you for your response.

However, I'm still a bit confused:Why does the CPU usage stay at 98% when using the hardware acceleration engine for hash algorithms, instead of decreasing?

0 Daniel 4 个月前回复 zhangxiaohui

TI__Mastermind 36245 points

Sa2ul driver doesn’t support sha offload from openssl. Team is working on a solution and will upstream it before releasing it as part of SDK.

Mean-while you can track the JIRA for the same: SITREQ-5764.

处理器

处理器论坛

PROCESSOR-SDK-AM62X: In AM62x SDK 10.01.04, the performance of the hardware acceleration engine is lower than that of the software.

Test Conclusions: AM62X SA2UL Hardware Acceleration Engine

Test Command

Test Data