A couple of weeks ago it was announced that Peer 1 Hosting had become the first company to create a CPU-GPU hybrid cloud that supports HPC system workloads. The service is built using Nvidia’s GPUs and comes loaded with the RealityServer application, an application that provides interactive and photorealistic applications remotely over the internet.
On the face of it, this sounds like a good idea and is part of the evolution of cloud for the HPC environment. Firstly, GPUs are good, proven performers in the computation stakes. See my other posts on GPUs. Second, using applications like RealityServer makes sense because it enables customers to interact with their data (or specifically models, 3D content in this situation) over the web.
However, for those customers not using web based applications like RealityServer, use of the cloud in the HPC environment can pose a problem. Firstly, do customers have the bandwidth available to take receipt of the data (which can be multiple Terabytes or Petabytes once processed)? Second, does the cloud service provider, which could have hundreds (potentially) of customers each generating multiple Terabytes of data, have the bandwidth available to send the data back to their customers? Finally, does the customer have the appropriate infrastructure to store the data created during the process?
These problems can be overcome – certainly integrators like OCF can provide adequate storage facilities for customers but none the less, customers must consider these questions before proceeding with a cloud service for HPC related problems.
[printprofile]
Loading ...
The questions about bandwidth for receiving the processed data is a valid question. One thing that may be overlooked though is compression of the processed data before transmission. Yes, the raw output of all the information processed by HPC solutions can be enormous. Same for the data to process. Generally speaking though, raw data to crunch can be very compressible.
If however you have video files to transcode, or data that is not easily compressible, a large circuit to transmit/receive the information would be necessary. Also, never underestimate the transfer speed of a USB drive shipped via FedEx if the source/destination sites are bandwidth constrained. Obviously, care should be taken to use encryption on sensitive data sent “off net”.
We have a lot of bandwidth in our DC’s, and can provide 10GbE connections if needed. Generally, bandwidth within our DC’s or out of our DC’s is not the issue. It’s the destination or customer’s side that can be the bottleneck in some cases.
So, leveraging a combination of compression, making sure the customer has enough bandwidth to transfer data from their side, and moving data outside of the net when necessary should be the approaches when deploying cloud based service.
Thanks for your comments!
Yes, raw data to crunch can be compressible but as you say, if the result is non-compressible or very large quantities of data then the bottleneck is generally the end-user!
In total agreement with the power of the ubiquitous USB drive! However, some of our customers just could not wait for an overnight FedEx shipment and require the data immediately after the computation has finished because of deadlines etc.
But certainly there are ways and means to overcome most obstacles as we are finding out in our Cloud business!
Hi Dave,
Whilst the utility of the USB drive is not questioned, what people often forget is how long it takes them to read! Most USB hard drives top out at 30MB/s which means that if you have 2Tb of data you are in for an 18 hour wait even after the postman man arrives. . .