question about QAT_HW and QAT_SW Co-existence

[gaozhangfei]

Hi

If both enable QAT_HW and QAT_SW, can it switch to SW when HW bandwidth resources is busy, and switch back to SW when HW resources is free again?

From https://github.com/intel/QAT_Engine/blob/master/docs/qat_coex.md For those algorithms that can achieve stronger performance with QAT_HW, the request will be offloaded to QAT_HW first, and after QAT_HW capacity is reached, it will be processed through QAT_SW. These algorithms include: RSA-2K/3K/4K, ECDSA-P384, ECDH-P256/P384/X25519, SM4-CBC(2048-16384 bytes).

Only find "after QAT_HW capacity is reached", is it mean get_instance, which is instances numbers instead of HW bandwidth.

(inst_num = get_instance(QAT_INSTANCE_ASYM, qat_svm)) == QAT_INVALID_INSTANCE) { WARN("Failed to get an instance\n"); if (qat_get_sw_fallback_enabled()) { CRYPTO_QAT_LOG("Failed to get an instance - fallback to SW - %s\n", func); *fallback = 1;

Can we switch from HW and SW according to HW bandwidth capability?

Thanks

[venkatesh6911]

Currently, the QAT_HW resource capacity is not being tracked. We rely on the status message CPA_STATUS_RETRY to do the switch to QAT_SW. This status message is returned when the requests are unable to process by the QAT_HW API at any point of time. https://github.com/intel/QAT_Engine/blob/b109a0a7678cca0a0a8ad0d35db519a13c9e88cc/qat_hw_rsa.c#L283 `sts = cpaCyRsaDecrypt(qat_instance_handles[inst_num], qat_rsaCallbackFn, &op_done, dec_op_data, output_buf); }

endif

    STOP_RDTSC(&qat_hw_rsa_dec_req_submit, 1, "[QAT HW RSA: submit]");
    if (sts == CPA_STATUS_RETRY) {
        DEBUG("cpaCyRsaDecrypt Retry \n");
        if (qat_rsa_coexist) {
            START_RDTSC(&qat_hw_rsa_dec_req_retry);
            ++num_rsa_priv_retry;
            qat_sw_rsa_priv_req += QAT_SW_SWITCH_MB8;
            *fallback = 1;
            qat_cleanup_op_done(&op_done);
            STOP_RDTSC(&qat_hw_rsa_dec_req_retry, 1, "[QAT HW RSA: retry]");
            return 0;
        }.....`

If the QAT_HW API does not return RETRY, then it will continue to process the requests with QAT_HW.

[gaozhangfei]

Thanks @venkatesh6911 for the clarification.

[gaozhangfei]

Hi, @venkatesh6911

Excuse me for more question to double check.

Does Intel qat hardware like Sapphire Rapids has the capability checking qat accelerator bandwidth?

The instance number does not mean bandwidth, right? For example, sometimes 1 instance may take lots of bandwidth. (inst_num = get_instance(QAT_INSTANCE_ASYM, qat_svm))

Any plan to consider checking accelerator bandwidth in qat engine?

Thanks

[venkatesh6911]

You are right. Instance number is not bandwidth. I assume that accelerator bandwidth meaning utilization for each QAT device ? No plan to include the utilization measurement in near future.

[gaozhangfei]

Thanks @venkatesh6911