AVX/MULX benchmarks

[Zstream]

Hi Everyone,

Does anyone know of any direct AVX benchmarks? I'm looking for AVX and MULX. Basically cryptographic performance.

 

[Cogman]

http://www.agner.org/optimize/instruction_tables.pdf

However, this is not basic cryptography performance. In order for the cryptography to go fast, it needs to use these instructions, which doesn't always (or often) happen in general encryption libraries.

In that case, you pretty much care about add, multiply, and rotation performance.

 

[Phynaz]

If it's crypto performance, what you are looking for are AES-NI benchmarks.

http://www.scottbrownconsulting.com...rformance-impact-of-hardware-accelerated-aes/
https://calomel.org/aesni_ssl_performance.html

 

[Zstream]

Well I'm trying to replicate https://software.intel.com/en-us/ar...-load-balancers-with-intel-xeon-v3-processors with multiple processors. It states that the software enables MULX instruction set? I haven't found a program to specifically task the CPU to get an understanding of HAProxy performance.

That is another test method without HAProxy.

 

[Phynaz]

Intel gives you the entire test setup, and it's all open source. Why can't you replicate it?

Or just script an openssl test.

 

[Zstream]

Intel gives you the entire test setup, and it's all open source. Why can't you replicate it?

Or just script an openssl test.

I want to test without using HAProxy as it can provide better statistical/reference points. Additionally, if someone has already benchmarked all of the cpus already, well more power to me

 

[Phynaz]

I'm confused, you're trying to replicate Intel's results without using the same benchmark? But the openssl benchmarks aren't what you want??

 

[Zstream]

I'm confused, you're trying to replicate Intel's results without using the same benchmark? But the openssl benchmarks aren't what you want??

What Cogman and you provided is great! I'm just looking for a way to determine if a bulldozer, athlon, haswell etc.. will naturally perform better. I'll be doing some benchmarks, but if there is a way to determine capability without going through a full purchase. These CPU's with different sockets add up fast!

 

[Phynaz]

Oh, OK, I believe Broadwell & Skylake have the latest crypto enhancements.
http://www.anandtech.com/show/10158/the-intel-xeon-e5-v4-review/3

You can also look for Truecrypt benchmarks.

 

[TheRyuu]

However, this is not basic cryptography performance. In order for the cryptography to go fast, it needs to use these instructions, which doesn't always (or often) happen in general encryption libraries.

Well provided you're using the latest OpenSSL/LibreSSL/BoringSSL library then you'll probably get them. I believe all of the above do a runtime selection of which version of a function to use. If your processor supports AVX2 then it'll use that version.

 

[Zstream]

I've been looking at the developer site and it looks like AES can be accelerated by the GPU using OpenCL. I'm not sure about Iris as of yet.

http://developer.amd.com/resources/...articles-whitepapers/bulk-encryption-on-gpus/

Has anyone seen benchmarks for AES acceleration using the GPU basically replacing FPGA boards?

 

[TheRyuu]

I've been looking at the developer site and it looks like AES can be accelerated by the GPU using OpenCL. I'm not sure about Iris as of yet.

http://developer.amd.com/resources/...articles-whitepapers/bulk-encryption-on-gpus/

Has anyone seen benchmarks for AES acceleration using the GPU basically replacing FPGA boards?

It may be useful but likely only for situations that are highly parallel. So whichever block cipher modes are parallelizable is probably where you may see the biggest gains if it turns out to be beneficial to run these on the GPU in the first place.

Stream ciphers like ChaCha could also probably benefit from GPU acceleration.