champtar / aestests Goto Github PK

# Compile
gcc -c -I${JAVA_HOME}/include/ -I${JAVA_HOME}/include/linux/ -fPIC -O2 *.c
gcc -O2 -o libjnitests.so -shared -Wl,-soname,libjnitests.so *.o -lcrypto
javac AESTests.java


# Run
java -Djava.library.path='./' AESTests
java -Djava.library.path='./' HMACSHA1Tests

# Tests
We encrypt 1000 - 20 bytes (20 is the offset)
We hash 1000 - 20 + 4 bytes (20 is the offset, 4 the roc) and write all 20 bytes of hmacsha1 at offset 1000

JAVA_CTR is a copy/paste of current libjitsi implementation, using JCE (AES/ECB/NoPadding) to encrypt each block (one call to cipher.doFinal() per block)

JCE_CTR is just using JCE CTR implementation (AES/CTR/NoPadding) (one call to cipher.init() and one to cipher.doFinal() per packet)
SunPKCS11_CTR is just using SunPKCS11 CTR implementation (AES/CTR/NoPadding) (one call to cipher.init() and one to cipher.doFinal() per packet)
BC_CTR is just using bouncy castle CTR implementation (AES/CTR/NoPadding) (one call to cipher.init() and one to cipher.doFinal() per packet)

OPENSSL_CTR is using OpenSSL CTR implementation (one JNI call per packet)


# Results (on my computer: core i7, fedora23)
OPENSSL_CTR : 396 ns
JAVA_CTR : 2610 ns
JCE_CTR : 3921 ns
BC_CTR : 8415 ns
SunPKCS11_CTR : 33747 ns

OPENSSL2_HMACSHA1 : 1503 ns
OPENSSL1_HMACSHA1 : 2214 ns
SunPKCS11_HMACSHA1 : 4153 ns
JCE_HMACSHA1 : 4417 ns
BC_HMACSHA1 : 5022 ns

OpenSSL speed (in 1000s of bytes per second)
openssl speed -evp AES-128-CTR
1024 bytes => 4618991.27k
openssl speed -evp SHA1
1024 bytes => 709489.46k


# Other benchs (on my computer)
A call to an no-op JNI function take ~7ns
GetByteArrayRegion take ~17ns (on 16 byte array)
GetPrimitiveArrayCritical + ReleasePrimitiveArrayCritical take ~25ns
so on the 390ns for an OpenSSL CTR round, 70ns (~20%) is data "transfer" (copy or pinning or ...)

# Further optimisation
We could do the Mac (hmac-sha1) in the same call