Super Micro Open Stroage Summit 2024

https://www.thecube.net/events/supermicro/open-storage-summit-2024

I attended the last 2 events at this stummit. Here are my thoughts and notes:

The New High Performance Computing: Optimized Storage from HPC to AI

Speakers:

• Randy Kreiser Supermicro
• CJ Newburn NVIDIA
• Balaji Venkateshwaran DDN
• Bill Panos Solidigm

CJ- The moderator looks like he is being held at gunpoint.

Rate at which usage models are changing

need for new infrastructure

new models characterized by large scale and making memory accesses of fine granularity

key-value store based storage systems?

At a rack level, there is much more integration. GPU is initiating requests from gpu, threads acting individually cpu/gpu direct storage.

Higher fine grained iops requirements from newer applications.

Balaji - DDN wants to be the one stop shop for storage and data management Streamline and optimize the entire flow for large scale datasystems Automate dataplacement and maximize flash performance. Secure multitenancy, maximize safe resource utilization.

As the utilization of data and blocks increases, are you able to maximize RoI of the other infrastructure? You want the ability to keep the GPU busy using an optimized end2end application stack.

Don’t want the storage to bottleneck the entire system.

Bill - Think about the subset of data types (block, object, files) needs to be collected/augmented on physical media. Needs to be optimized against density per terrabyte. High Capacity and sequential writes. Sequential writes happen in real time. You want a product that can do ingress and egress at the same level of bandwidth for AI applications.

Randy - Liquid cooling is becoming an important thing. Musk is designing his new datacentres around liquid cooling. A lot of what we talked about has requirements like 1.8tb/s use pcie gen5 right now, pice gen6 in the future. You’re always going to need hard drives. We’re getting to exascale systems. Nvme is getting to a terrabyte per second speeds. We’re looking at virtually 10tb/s in the near future (provided by ddn).

CJ - GPU use cases have been changing from compute monsters to data io monsters. thousands of threads doing individual accesses Accessing storage that might not fit in memory in concurrently, which means you need high bandwidth pcie. working scaled accelerated access (scata).

Drastically increase the iops, and do fine grained access and do scat and gather and pass data through pcie busses as fast as possible.

Another big concern is security. The compute infra is so large, we can’t afford to not share it. Storage is a vulnerability. Partnering with DDN on infinea. Plugin something at user level that sends file requests to the dpu on pcie bus which should drastically improve security

Bill - Gen5 just started shipping in volume, wild to be thinking about Gen6 already. The progression in interface is speeding up…felt like we were on Gen3 for a lifetime.

Balaji -

the datapath from flash to the gpu is the thread that ties these companies together. As the use cases and data types explode, you want to abstract away complexity for the customers.

The entire datapath between the gpu and flash - is secure scalable and fast.

Lots of coprocessors coming to meet future applications and optimize for power and performance.