GPU Profiling & Performance Optimization - The industry emphasizes improving performance and saving costs by optimizing software, potentially reducing server usage by 10% [4] - Sampled profiling is used to balance data volume and continuous monitoring, with examples of sampling 100 times per second resulting in less than 1% CPU overhead and 4MB memory overhead [5] - The industry highlights the importance of production environment profiling to observe real-world application performance with low overhead [8] - The company's solution leverages Linux EVPF, enabling profiling without application instrumentation [9] Technology & Metrics - The company's GPU profiling solution uses Nvidia NVML to extract metrics, including overall node utilization (blue line), individual process utilization (orange line), memory utilization, and clock speed [11][12] - Key metrics include power utilization (with power limit as a dashed line), temperature (important to avoid throttling at 80 degrees Celsius), and PCIe throughput (negative for receiving, positive for sending, e g 10 MB/s) [13][14] - The solution correlates GPU metrics with CPU profiles collected using EVPF to analyze CPU activity during periods of less than full GPU utilization [14] GPU Time Profiling - The company introduces GPU time profiling to measure time spent on individual CUDA functions, determining start and end times of kernels via the Linux kernel [18] - The solution displays CPU stacks with leaf nodes representing functions taking time on the GPU, with colors indicating different binaries (e g blue for Python) [19][20] Deployment & Integration - The company's solution can be deployed using a binary on Linux, Docker, or as a DaemonSet on Kubernetes, requiring a manifest YAML and token [21] - Turbo Puffer is interested in integrating the company's GPU profiling to improve the performance of their vector engine [22]