Puget Systems offers cost effective DIGITS solution vs. the cloud
About RentPath
RentPath is a leading digital marketing solutions company that empowers millions nationwide to find apartments and houses for rent. Through its brands (ApartmentGuide.com, Rent.com, Rentals.com, LiveLovely.com), RentPath continues to simplify the apartment search experience while driving quality advertiser leads that result in increased occupancy rates and a high return on investment. With powerful online and mobile solutions that provide prospective renters with the information and tools they need, RentPath connects consumers with homes that reflect their personal lifestyles.
RentPath helps people navigate the rental journey by providing a delightful, stress-free experience so that all renters can find and enjoy their ideal home.
Tools of the Trade
The RentPath data science team uses the two gpu boxes for a variety of deep learning tasks. Examples of projects based on gpu-accelerated computations include, but are not limited to, image recognition (caffe; NLP tasks, such as part of speech tagging and named entity recognition using BI-LSTMs in tensorflow); word2vec implementation to support RentPath’s recommendation engine and many more.
Challenges
One of the RentPath data science team’s first tasks was to do image classification with million of images showing apartment interiors and outdoor spaces. This is a time and resource-intensive machine learning application. They explored using a cloud service and the most powerful gpu instance at the time had a cost of around $14 an hour. Training the model took about two weeks to complete, at a cost of about $4800.
In order to improve the model performance, retraining it would be necessary, with no guarantee that the new method will outperform the prior one. Every question carried a high expense.
As David McFarland, Head of Data Science at RentPath, explains, “I didn’t want cost considerations to overly impact my team’s decision to run a model.”
Puget Systems Solution
We put together a solution based off our recommended system for Nvidia DIGITS. It is a pair of workstations specifically designed for Deep Learning workloads like DIGITS with Caffe. The systems each have 4x NVIDIA GeForce GTX 1080 Ti GPUs, all running at x16 speed making it a very GPU-focused workstation housed in an chassis that has optimal cooling and quiet operations. The setup also had built in dual 10GB ports for faster network data transfers, 128 GB of ram, a large SSD and a massive hard drive.
Performance-wise, each machine was three to fours times faster than what they were getting from the cloud and paid for itself in saved cloud costs after one month of 100% utilization.
Why RentPath Chose Puget Systems
Puget Systems built a box specifically for what the RentPath data science team was doing. Puget understood what was being done, installed software that RentPath was using, and provided adequate testing using the same tools RentPath was going to use.
David remarked, “Puget Systems has a great understanding of what this system is going be used for and built it with that in mind. Technicians even installed and tested it with CUDA to make sure it performed as expected. Having our own hardware has simplified our decision-making around model-testing. If we want to run a model, we just do it. We appreciate the freedom the solution has provided.”
