I’m a huge fan of Joyent, and the SmartDataCenter (now Triton) platform is one that I would love to use for machine learning and data science. The emphasis on observability and post-mortem debugging makes it a fantastic platform for production systems and by virtue of being totally open source (unlike, say, AWS) it is impossible to be truly locked into the platform. It is also usable for on-premise development environments, which can result in huge savings versus testing and developing in a public cloud.
I am envisioning a series of Docker containers dishing out inference requests to TensorFlow Serving models which were developed and trained in an on-premise Triton deployment. I began work on Passer to facilitate this design, and while it would work fine on CPUs, accelerating training and inference with GPUs would be ideal for peak performance.
There are two issues with this.
If PCI passthrough worked, we could provision KVM Linux instances through Triton and just use the native Linux NVidia drivers. This would probably be less of an uphill battle than trying to convince NVidia to do anything.
So, what are the issues with PCI passthrough? I believe that all of the systems at Joyent engineering have VT-d support, so we’re essentially facing driver issues. In particular…
The IOMMU interfaces are present, but the AMD64 code will need a rewrite (according to @rmustacc). I’m attributing that just to be clear that I wasn’t making any value judgements here, but the point is that this isn’t an infeasible thing.
<rmustacc> Like anything else at that layer, it's not particularly hard as it is detailed. <rmustacc> You need to be very careful not to overlook something.
So, I may get around to exploring this in more depth at some point.
For now, I’m just making some notes for posterity.Written on March 27th, 2016 by Alex Caudill