Message ID | 20191017145654.11371-3-berrange@redhat.com |
---|---|
State | New |
Headers | show |
Series | crypto: improve performance of ciphers in XTS mode | expand |
On 10/17/19 4:56 PM, Daniel P. Berrangé wrote: > Currently the crypto benchmarks are processing data in varying chunk > sizes, over a fixed time period. This turns out to be a terrible idea > because with small chunk sizes the overhead of checking the elapsed > time on each loop iteration masks the true performance. > > Benchmarking over a fixed data size avoids the loop running any system > calls which can interfere with the performance measurements. > > Before this change > > Enc chunk 512 bytes 2283.47 MB/sec Dec chunk 512 bytes 2236.23 MB/sec OK > Enc chunk 4096 bytes 2744.97 MB/sec Dec chunk 4096 bytes 2614.71 MB/sec OK > Enc chunk 16384 bytes 2777.53 MB/sec Dec chunk 16384 bytes 2678.44 MB/sec OK > Enc chunk 65536 bytes 2809.34 MB/sec Dec chunk 65536 bytes 2699.47 MB/sec OK > > After this change > > Enc chunk 512 bytes 2058.22 MB/sec Dec chunk 512 bytes 2030.11 MB/sec OK > Enc chunk 4096 bytes 2699.27 MB/sec Dec chunk 4096 bytes 2573.78 MB/sec OK > Enc chunk 16384 bytes 2748.52 MB/sec Dec chunk 16384 bytes 2653.76 MB/sec OK > Enc chunk 65536 bytes 2814.08 MB/sec Dec chunk 65536 bytes 2712.74 MB/sec OK > > The actual crypto performance hasn't changed, which shows how > significant the mis-measurement has been for small data sizes. > > Signed-off-by: Daniel P. Berrangé <berrange@redhat.com> > --- > tests/benchmark-crypto-cipher.c | 26 ++++++++++++++------------ > tests/benchmark-crypto-hash.c | 17 +++++++++-------- > 2 files changed, 23 insertions(+), 20 deletions(-) > > diff --git a/tests/benchmark-crypto-cipher.c b/tests/benchmark-crypto-cipher.c > index 3ca31a2779..d8db5504d4 100644 > --- a/tests/benchmark-crypto-cipher.c > +++ b/tests/benchmark-crypto-cipher.c > @@ -21,11 +21,12 @@ static void test_cipher_speed(size_t chunk_size, > { > QCryptoCipher *cipher; > Error *err = NULL; > - double total = 0.0; > uint8_t *key = NULL, *iv = NULL; > uint8_t *plaintext = NULL, *ciphertext = NULL; > size_t nkey; > size_t niv; > + size_t total = 2 * GiB; Can you use 'const size_t total'?. > + size_t remain; > > if (!qcrypto_cipher_supports(alg, mode)) { > return; > @@ -58,33 +59,34 @@ static void test_cipher_speed(size_t chunk_size, > &err) == 0); > > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > g_assert(qcrypto_cipher_encrypt(cipher, > plaintext, > ciphertext, > chunk_size, > &err) == 0); > - total += chunk_size; > - } while (g_test_timer_elapsed() < 1.0); > + remain -= chunk_size; > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("Enc chunk %zu bytes ", chunk_size); > - g_print("%.2f MB/sec ", total / g_test_timer_last()); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > - total = 0.0; > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > g_assert(qcrypto_cipher_decrypt(cipher, > plaintext, > ciphertext, > chunk_size, > &err) == 0); > - total += chunk_size; > - } while (g_test_timer_elapsed() < 1.0); > + remain -= chunk_size; > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("Dec chunk %zu bytes ", chunk_size); > - g_print("%.2f MB/sec ", total / g_test_timer_last()); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > qcrypto_cipher_free(cipher); > g_free(plaintext); > diff --git a/tests/benchmark-crypto-hash.c b/tests/benchmark-crypto-hash.c > index 9b6f7a9155..67b7067223 100644 > --- a/tests/benchmark-crypto-hash.c > +++ b/tests/benchmark-crypto-hash.c > @@ -20,7 +20,8 @@ static void test_hash_speed(const void *opaque) > size_t chunk_size = (size_t)opaque; > uint8_t *in = NULL, *out = NULL; > size_t out_len = 0; > - double total = 0.0; > + size_t total = 2 * GiB; Ditto: 'const size_t total'. > + size_t remain; > struct iovec iov; > int ret; > > @@ -31,20 +32,20 @@ static void test_hash_speed(const void *opaque) > iov.iov_len = chunk_size; > > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > ret = qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256, > &iov, 1, &out, &out_len, > NULL); > g_assert(ret == 0); > > - total += chunk_size; > - } while (g_test_timer_elapsed() < 5.0); > + remain -= chunk_size; > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("sha256: "); > - g_print("Testing chunk_size %zu bytes ", chunk_size); > - g_print("done: %.2f MB in %.2f secs: ", total, g_test_timer_last()); > - g_print("%.2f MB/sec\n", total / g_test_timer_last()); > + g_print("Hash %zu GB chunk size %zu bytes ", total / GiB, chunk_size); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > g_free(out); > g_free(in); > Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
On Thu, Oct 17, 2019 at 03:56:52PM +0100, Daniel P. Berrangé wrote: > Currently the crypto benchmarks are processing data in varying chunk > sizes, over a fixed time period. This turns out to be a terrible idea > because with small chunk sizes the overhead of checking the elapsed > time on each loop iteration masks the true performance. > > Benchmarking over a fixed data size avoids the loop running any system > calls which can interfere with the performance measurements. > > Before this change > > Enc chunk 512 bytes 2283.47 MB/sec Dec chunk 512 bytes 2236.23 MB/sec OK > Enc chunk 4096 bytes 2744.97 MB/sec Dec chunk 4096 bytes 2614.71 MB/sec OK > Enc chunk 16384 bytes 2777.53 MB/sec Dec chunk 16384 bytes 2678.44 MB/sec OK > Enc chunk 65536 bytes 2809.34 MB/sec Dec chunk 65536 bytes 2699.47 MB/sec OK > > After this change > > Enc chunk 512 bytes 2058.22 MB/sec Dec chunk 512 bytes 2030.11 MB/sec OK > Enc chunk 4096 bytes 2699.27 MB/sec Dec chunk 4096 bytes 2573.78 MB/sec OK > Enc chunk 16384 bytes 2748.52 MB/sec Dec chunk 16384 bytes 2653.76 MB/sec OK > Enc chunk 65536 bytes 2814.08 MB/sec Dec chunk 65536 bytes 2712.74 MB/sec OK > > The actual crypto performance hasn't changed, which shows how > significant the mis-measurement has been for small data sizes. > > Signed-off-by: Daniel P. Berrangé <berrange@redhat.com> > --- > tests/benchmark-crypto-cipher.c | 26 ++++++++++++++------------ > tests/benchmark-crypto-hash.c | 17 +++++++++-------- > 2 files changed, 23 insertions(+), 20 deletions(-) > > diff --git a/tests/benchmark-crypto-cipher.c b/tests/benchmark-crypto-cipher.c > index 3ca31a2779..d8db5504d4 100644 > --- a/tests/benchmark-crypto-cipher.c > +++ b/tests/benchmark-crypto-cipher.c > @@ -21,11 +21,12 @@ static void test_cipher_speed(size_t chunk_size, > { > QCryptoCipher *cipher; > Error *err = NULL; > - double total = 0.0; > uint8_t *key = NULL, *iv = NULL; > uint8_t *plaintext = NULL, *ciphertext = NULL; > size_t nkey; > size_t niv; > + size_t total = 2 * GiB; > + size_t remain; > > if (!qcrypto_cipher_supports(alg, mode)) { > return; > @@ -58,33 +59,34 @@ static void test_cipher_speed(size_t chunk_size, > &err) == 0); > > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > g_assert(qcrypto_cipher_encrypt(cipher, > plaintext, > ciphertext, > chunk_size, > &err) == 0); > - total += chunk_size; > - } while (g_test_timer_elapsed() < 1.0); > + remain -= chunk_size; > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("Enc chunk %zu bytes ", chunk_size); > - g_print("%.2f MB/sec ", total / g_test_timer_last()); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > - total = 0.0; > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > g_assert(qcrypto_cipher_decrypt(cipher, > plaintext, > ciphertext, > chunk_size, > &err) == 0); > - total += chunk_size; > - } while (g_test_timer_elapsed() < 1.0); > + remain -= chunk_size; Are we sure that total is a multiple of chunk_size? Maybe I would have increased 'done' and checked (done <= total), but maybe it doesn't matter since I think that chunk_size will always be a power of two. > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("Dec chunk %zu bytes ", chunk_size); > - g_print("%.2f MB/sec ", total / g_test_timer_last()); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > qcrypto_cipher_free(cipher); > g_free(plaintext); > diff --git a/tests/benchmark-crypto-hash.c b/tests/benchmark-crypto-hash.c > index 9b6f7a9155..67b7067223 100644 > --- a/tests/benchmark-crypto-hash.c > +++ b/tests/benchmark-crypto-hash.c > @@ -20,7 +20,8 @@ static void test_hash_speed(const void *opaque) > size_t chunk_size = (size_t)opaque; > uint8_t *in = NULL, *out = NULL; > size_t out_len = 0; > - double total = 0.0; > + size_t total = 2 * GiB; > + size_t remain; > struct iovec iov; > int ret; > > @@ -31,20 +32,20 @@ static void test_hash_speed(const void *opaque) > iov.iov_len = chunk_size; > > g_test_timer_start(); > - do { > + remain = total; > + while (remain) { > ret = qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256, > &iov, 1, &out, &out_len, > NULL); > g_assert(ret == 0); > > - total += chunk_size; > - } while (g_test_timer_elapsed() < 5.0); > + remain -= chunk_size; > + } > + g_test_timer_elapsed(); > > - total /= MiB; > g_print("sha256: "); > - g_print("Testing chunk_size %zu bytes ", chunk_size); > - g_print("done: %.2f MB in %.2f secs: ", total, g_test_timer_last()); > - g_print("%.2f MB/sec\n", total / g_test_timer_last()); > + g_print("Hash %zu GB chunk size %zu bytes ", total / GiB, chunk_size); > + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); > > g_free(out); > g_free(in); Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
diff --git a/tests/benchmark-crypto-cipher.c b/tests/benchmark-crypto-cipher.c index 3ca31a2779..d8db5504d4 100644 --- a/tests/benchmark-crypto-cipher.c +++ b/tests/benchmark-crypto-cipher.c @@ -21,11 +21,12 @@ static void test_cipher_speed(size_t chunk_size, { QCryptoCipher *cipher; Error *err = NULL; - double total = 0.0; uint8_t *key = NULL, *iv = NULL; uint8_t *plaintext = NULL, *ciphertext = NULL; size_t nkey; size_t niv; + size_t total = 2 * GiB; + size_t remain; if (!qcrypto_cipher_supports(alg, mode)) { return; @@ -58,33 +59,34 @@ static void test_cipher_speed(size_t chunk_size, &err) == 0); g_test_timer_start(); - do { + remain = total; + while (remain) { g_assert(qcrypto_cipher_encrypt(cipher, plaintext, ciphertext, chunk_size, &err) == 0); - total += chunk_size; - } while (g_test_timer_elapsed() < 1.0); + remain -= chunk_size; + } + g_test_timer_elapsed(); - total /= MiB; g_print("Enc chunk %zu bytes ", chunk_size); - g_print("%.2f MB/sec ", total / g_test_timer_last()); + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); - total = 0.0; g_test_timer_start(); - do { + remain = total; + while (remain) { g_assert(qcrypto_cipher_decrypt(cipher, plaintext, ciphertext, chunk_size, &err) == 0); - total += chunk_size; - } while (g_test_timer_elapsed() < 1.0); + remain -= chunk_size; + } + g_test_timer_elapsed(); - total /= MiB; g_print("Dec chunk %zu bytes ", chunk_size); - g_print("%.2f MB/sec ", total / g_test_timer_last()); + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); qcrypto_cipher_free(cipher); g_free(plaintext); diff --git a/tests/benchmark-crypto-hash.c b/tests/benchmark-crypto-hash.c index 9b6f7a9155..67b7067223 100644 --- a/tests/benchmark-crypto-hash.c +++ b/tests/benchmark-crypto-hash.c @@ -20,7 +20,8 @@ static void test_hash_speed(const void *opaque) size_t chunk_size = (size_t)opaque; uint8_t *in = NULL, *out = NULL; size_t out_len = 0; - double total = 0.0; + size_t total = 2 * GiB; + size_t remain; struct iovec iov; int ret; @@ -31,20 +32,20 @@ static void test_hash_speed(const void *opaque) iov.iov_len = chunk_size; g_test_timer_start(); - do { + remain = total; + while (remain) { ret = qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256, &iov, 1, &out, &out_len, NULL); g_assert(ret == 0); - total += chunk_size; - } while (g_test_timer_elapsed() < 5.0); + remain -= chunk_size; + } + g_test_timer_elapsed(); - total /= MiB; g_print("sha256: "); - g_print("Testing chunk_size %zu bytes ", chunk_size); - g_print("done: %.2f MB in %.2f secs: ", total, g_test_timer_last()); - g_print("%.2f MB/sec\n", total / g_test_timer_last()); + g_print("Hash %zu GB chunk size %zu bytes ", total / GiB, chunk_size); + g_print("%.2f MB/sec ", (double)total / MiB / g_test_timer_last()); g_free(out); g_free(in);
Currently the crypto benchmarks are processing data in varying chunk sizes, over a fixed time period. This turns out to be a terrible idea because with small chunk sizes the overhead of checking the elapsed time on each loop iteration masks the true performance. Benchmarking over a fixed data size avoids the loop running any system calls which can interfere with the performance measurements. Before this change Enc chunk 512 bytes 2283.47 MB/sec Dec chunk 512 bytes 2236.23 MB/sec OK Enc chunk 4096 bytes 2744.97 MB/sec Dec chunk 4096 bytes 2614.71 MB/sec OK Enc chunk 16384 bytes 2777.53 MB/sec Dec chunk 16384 bytes 2678.44 MB/sec OK Enc chunk 65536 bytes 2809.34 MB/sec Dec chunk 65536 bytes 2699.47 MB/sec OK After this change Enc chunk 512 bytes 2058.22 MB/sec Dec chunk 512 bytes 2030.11 MB/sec OK Enc chunk 4096 bytes 2699.27 MB/sec Dec chunk 4096 bytes 2573.78 MB/sec OK Enc chunk 16384 bytes 2748.52 MB/sec Dec chunk 16384 bytes 2653.76 MB/sec OK Enc chunk 65536 bytes 2814.08 MB/sec Dec chunk 65536 bytes 2712.74 MB/sec OK The actual crypto performance hasn't changed, which shows how significant the mis-measurement has been for small data sizes. Signed-off-by: Daniel P. Berrangé <berrange@redhat.com> --- tests/benchmark-crypto-cipher.c | 26 ++++++++++++++------------ tests/benchmark-crypto-hash.c | 17 +++++++++-------- 2 files changed, 23 insertions(+), 20 deletions(-)