Crypto performance
Some brief experimentation with crypto performance on more traditional processors:
These numbers where obtained running the OpenSSL (0.9.8e) speed test, first using a single core, then using all of the cores.
Clovertown (Dell PowerEdge 2900) (2.66GHz) Single thread ============= rsa-1024 (sign) 1200 Ops/sec aes-128-cbc 135MB/s Maximum Throughput for single-processor (4-cores) ================================================= rsa-1024 (sign) 4750 Ops/sec aes-128-cbc 525MB/s [Many thanks to Chi-Chang for collecting this data]
Now, it is probably possible to improve on these numbers. Certainly, OpenSSL may not contain the optimal RSA or AES implementation for Clovertown. However, it does give a ballpark estimate for the crypto performance that can be delivered by traditional processors, even those with Stellar single-thread performance.
Further, given this crypto processing is performed in SW, in order to achieve this level of performance, each and every cycle on each and every core is consumed performing the crypto processing, leaving no idle cycles to do anything meaningful with the data being generated.....
If we contrast this with the peak performance that can be delivered by the T2 crypto accelerators (see earlier post), it is apparent that there is a significant upside to supporting HW crypto accelerators, if you are interested in secure application performance. Further, with when using the T2 accelerators, because the crypto processing is offloaded to the HW, the cores still have idle cycles, which can be used to process the data that is being produced!
