System-wide Effects

To determine the system-wide benefits of offloading cryptographic processing, we run multiple threads (up to 24) of the openssl speed benchmark with various algorithms, while at the same time we run a simple CPU-intensive job. The CPU ``hog'' process consists of a small program that performs 2³² function calls, each function call performing an integer-multiply operation. The elapsed time for the CPU hog process was recorded for each (algorithm, number of threads) tuple. As we see in Figure 4, the crypto accelerators very effectively eliminate contention for the otherwise-shared resource, the CPU, whether the crypto performed is symmetric (DES, 3DES) or asymmetric (DSA with 1024-bit keys). The execution time for the hog process remains constant, regardless of the number of threads of execution.

  

Figure 4: Program execution time while multiple threads perform crypto-operations in parallel. The bars show the elapsed time in seconds for executing the CPU-bound process for different algorithms and numbers of threads.