Every AES encrypt/decrypt call is doing a heap allocation/free about gamenetworkingsockets HOT 7 CLOSED

Actually, it's doing multiple allocations.

I checked in a custom allocator that will put them into an arena. It made the CBC encryption about 10% faster on a Ubuntu VM I'm using. We'll see what numbers @tycho comes up with for AES-GCM, on the allocation sizes we care about most. If it's not a decent win, maybe back it out or make it an option? Still, regardless of the numbers, it really bothers me to be making multiple heap calls per packet!

Maybe the heap is fast nowadays and I'm just a dinosaur.

from gamenetworkingsockets.

^ For reference: 17c52cb

from gamenetworkingsockets.

The numbers are in!

I am a dinosaur.

I'll back out that "optimization". Closing this.

from gamenetworkingsockets.

I think the perf diff would depend on the heap implementation. Most modern implementations use a thread-local small allocation pool.

from gamenetworkingsockets.

FWIW Unreal isn't doing anything special around EVP_CIPHER_CTX_new,

from gamenetworkingsockets.

Unreal uses jmalloc for the whole engine though.

from gamenetworkingsockets.

Re-opening this issue. Just so I can officially close it for really realz.

from gamenetworkingsockets.