(Virtual) Instrumenting your hot path in C++

Description

At Thunder Compute we build GPU virtualization software which uses a user space shim library to intercept CUDA calls and send them over the network. AI applications make thousands of CUDA calls per second, so we’ve built custom observability to collect accurate logs while minimizing performance overhead. We’ll talk about some of the novel techniques we use and the fun journey to optimize this system.

Details

The presentation will begin at 5:30
This event is for virtual attendance. For in-person attendance, please see the alternative page.

When: Tuesday, August 4th
Presentation begins at 5:30 pm

This event has been graciously sponsored by our good friends at Selby Jennings