On 12/07/2015 02:43 PM, Software Engineer 979 wrote: > Hello, > > I'm currently developing an data transfer application using OpenSSL. > The application is required to securely transfer large amounts of data > over a low latency/high bandwidth network. The data being transferred > lives in a 3rd part application that uses 1 MB buffer to transfer data > to my application. When I hook OpenSSL into my application I notice an > appreciable decline in network throughput. I've traced the issue the > default TLS record size of 16K. The smaller record size causes the 3rd > party application's buffer to be segmented into 4 16K buffers per > write and the resulting overhead considerably slows things down. I've > since modified the version of OpenSSL that I'm using to support an > arbitrary TLS record size allowing OpenSSL to scale up to 1MB or > larger TLS record size. Since this change, my network throughput has > dramatically increased (187% degradation down to 33%). > > I subsequently checked the TLS RFC to determine why a 16K record size > was being used, and all could find was the following: > > length > The length (in bytes) of the following TLSCompressed.fragment. > The length MUST NOT exceed 2^14 + 1024. > > The language here is pretty explicit stating that the length must not > exceed 16K (+ some change).Does anyone know the reason for this? Is > there a cryptographic reason why we shouldn't exceed this message > size? Based on my limited experiment, it would appear that a larger > record size would benefit low latency/high bandwidth networks. > The peer is required to buffer the entire record before processing it, at at that point the data could be from an untrusted party/attacker. So the limit is for protection against denial-of-service via resource exhaustion. -Ben -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://mta.openssl.org/pipermail/openssl-users/attachments/20151207/5f27400b/attachment.html>
