IBM and CoreWeave combat agent workload isolation

Isolating agent workloads becomes a practical infrastructure challenge as research teams move beyond training models to running code and testing agents. Systems designed for intensive computing power now need to handle workloads that interact with tools, storage, and other services.

IBM Research’s infrastructure must support increasingly diverse workloads as model development practices evolve. According to Brian Belgodere (pictured), senior technical officer at International Business Machines Corp., reinforcement learning adds a task execution phase that changes what these systems need to do.

“The RL process is [that] You’re in the process of training a model,” he said. “At some point, you take that control point and then you actually load it into inference, ask it to do something, and you measure.” This is your testing period.”

Belgodere spoke with theCUBE Research’s Dave Vellante and John Furrier at the Fully Connected event during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed evolving research workloads, security-performance tradeoffs, shared infrastructure engineering, and agent workload isolation. (*Disclosure below.)

From model training to agent code execution

IBM Research’s work on its Granite family of models required extensive computing infrastructure. According to Belgodere, the cooling and power requirements of a subsequent generation of hardware were key to IBM’s decision to work with CoreWeave Inc.

“That was the trigger for me: ‘Hey, build out our infrastructure,'” he said. “We went out and built a large H100 cluster, and we did it ourselves: we got the space, right down to the smallest detail. It was a huge task.”

The relationship has expanded to include collaborative engineering around identity management and workload controls. According to Belgodere, IBM provided requirements for extending its internal identity systems in CoreWeave and then refined the implementation in several iterations.

“A lot of times they come to us and say, ‘Hey, we’re thinking about it, can we get your feedback?'” he said. “And we’ll be happy to get through it.”

Agent workload isolation becomes a design requirement

A large portion of the IBM Research cluster is a single-tenant cluster with its own storage provisioned within CoreWeave and additional capacity available within cost and security parameters, Belgodere said. IBM’s collaboration with CoreWeave also includes CoreWeave sandboxes, which support isolated execution on dedicated infrastructure or through a managed serverless runtime. These options allow researchers to choose where the agent code runs and what resources it can access.

“There are a lot of mistakes and people tend to underestimate the cost of changing some of these decisions,” Belgodere said. “If you make a bad architectural decision early on, either the cost will be high [that] You accidentally bought way too much network infrastructure, or you have to buy and retool everything.”

IBM measures the performance impact of security controls using benchmark results and uses these results to discuss tradeoffs with security teams. Workload isolation, along with corporate identity integration, is part of this broader security architecture.

“This is a supply chain problem from top to bottom,” he said. “It’s not just about the hardware, but also the firmware, the kernel levels, the code, your data lineage. Then you get into the whole world of agents, your images. It’s an absolute lineage problem.”

Here is the full video interview, part of SiliconANGLE and theCUBE’s coverage of the Fully Connected event:

(*Disclosure: TheCUBE is a paid media partner for the Fully Connected event. Neither CoreWeave, the sponsor of theCUBE’s coverage, nor other sponsors have editorial control over theCUBE or SiliconANGLE content.)

Photo: SiliconANGLE