Cracking hashes using AWS GPU instances

The Concept

Use AWS to run a timeboxed hashcat instance with many GPU's to test the strength of password/key hashes.

The intended architecture is:

• use slack to trigger a Lambda function (via standard bot-http interface)
• the Lambda function would instance.start() an ec2 multi GPU instance
• which on start up would configure hashcat using configurations and dictionaries off an S3 bucket
• let hashcat run for a set period
• return the result via a http webhook to slack.

EC2 Instance

I set up an ec2 instance as a proof of concept to run hashcat. AWS provides an instance with 4, 8 or 16 GPU's; I chose the g3.4xlarge. For the OS, I'm using the NVIDIA Linux AMI which comes with the NVIDIA drivers ready to go. This is important, as otherwise you need to modify kernel drivers which requires rebooting the system. I just want something easy.

EC2 user-data

I'm using user-data as part of the ec2 config to set a couple of parameters, install hashcat and run the benchmarks. In the real version, this would run a script to retrieve hashcat configs and dictionaries from S3 as well as the user supplied hash to be cracked.

#!/bin/sh
sudo nvidia-smi -pm 1
sudo nvidia-smi -acp 0
sudo nvidia-smi —auto-boost-default=1
git clone https://github.com/hashcat/hashcat.git
cd hashcat/
git submodule update --init
make
sudo make install
/usr/local/bin/hashcat -b --force >> /var/log/cloud-init-output.log
halt

Results

Disappointing! While this all works in theory, the GPU's afforded through AWS are no where near the power of a current generation gaming GPU. It was much quicker to run this on my desktop than in the cloud - even across 4 GPU's. It should also be noted that the configuration and availability of the high end compute instances varies significantly.

So I left it here. I may pick it up again if/when I find a way to do this on AWS that beats a mid-range GPU.

Hash rates on AWS

OpenCL Platform #1: NVIDIA Corporation
======================================
* Device #1: GRID K520, 1023/4095 MB allocatable, 8MCU
* Device #2: GRID K520, 1023/4095 MB allocatable, 8MCU
* Device #3: GRID K520, 1023/4095 MB allocatable, 8MCU
* Device #4: GRID K520, 1023/4095 MB allocatable, 8MCU

Hashtype: MD5

Speed.Dev.#1.....:  2357.8 MH/s (56.86ms)
Speed.Dev.#2.....:  2359.7 MH/s (56.85ms)
Speed.Dev.#3.....:  2357.9 MH/s (56.89ms)
Speed.Dev.#4.....:  2353.3 MH/s (56.99ms)
Speed.Dev.#*.....:  9428.8 MH/s

Hashtype: SHA1

Speed.Dev.#1.....:   629.8 MH/s (53.24ms)
Speed.Dev.#2.....:   630.0 MH/s (53.22ms)
Speed.Dev.#3.....:   632.0 MH/s (53.06ms)
Speed.Dev.#4.....:   629.2 MH/s (53.29ms)
Speed.Dev.#*.....:  2521.0 MH/s

Hash rates on GTX 1060

OpenCL Platform #1: NVIDIA Corporation
======================================
* Device #1: GeForce GTX 1060 6GB, 1536/6144 MB allocatable, 10MCU

Hashtype: MD5

Speed.Dev.#1.....: 11437.2 MH/s (58.60ms)

Hashtype: SHA1

Speed.Dev.#1.....:  4310.1 MH/s (77.76ms)