It's nice to see the comparison of the AES performance with and without the AES-NI support. I had the impression from other threads that the perfromance was 3x, so I was surprised to see your report of >7x.
It should be noted that presented number of ~770MBps for 128-bit key is for CBC encrypt. CBC decrypt is 3x faster than encrypt. Counter mode would also be 3x faster than presented number.
One thing that struck me as womehat odd was that the 'before' numbers seemed a bit low. Unfortunately, you don't indicate any details of the system you tested on; in particular the clock-rate.
I also wonder why one doesn't see ~2x increase from 256 to 1024 bytes chunk. There is logic that prevents it on hyper-threading CPUs with shared L1 cache (for security reasons), but P4 should be the only one "suffering." Atom might do too...
For comparison, on a 3.0 GHz Xeon, I get the following numbers with v1.0.0: type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes aes-128 cbc 87700.71k 95696.04k 98280.53k 207097.86k 210802.01k
Meaning that large chunks are processed in ~15 cycles per byte processed with 128-bit key, which is expected. AES-NI CBC encrypt runs in ~4 cycles per byte and decrypt in ~1.3. In other words AES-NI CBC encrypt is commonly ~3.75 times faster than software and decrypt ~11x. A.
______________________________________________________________________ OpenSSL Project http://www.openssl.org Development Mailing List [email protected] Automated List Manager [email protected]
