** Special thanks to my classmates who were part of this project. **
The Advanced Computing Center for Research and Education (ACCRE) is a computer cluster serving the high-performance computing needs of research for Vanderbilt University. In this data question, you will be analyzing data on jobs run on ACCRE's hardware.
The computing resources in the ACCRE cluster are divided into nodes. Each node is equipped with some number of CPUs, and some of the nodes contain GPUs (graphical processing units). You have been provided a dataset, accre-gpu-jobs-2022.csv, which contains information on jobs submitted to ACCRE's GPU partitions. Traditionally GPUs were designed to power video games to perform calculations quickly. Because of the nature of their design, GPUs are being used more and more for non-graphics applications as well (e.g. for deep learning applications, molecular dynamics, image processing, and much more).
Recently, more researchers are performing work that requires GPU acceleration. Your task in this project is to analyze the GPU jobs that have been run on ACCRE over the last few years in order to better understand what needs to be provisioned moving forward.
ACCRE has three scheduler partitions that groups can get access to which consist of machines each with 4 GPU cards. Each partition is named after the nvidia processor generation that it contains, from oldest to newest: maxwell, pascal, and turing. GPU nodes are requested per-GPU, and users can request up to 4 GPUs per node. For each GPU the user is allocated up to one quarter of the machine's RAM and CPU cores. All of the current GPU-accelerated hardware is also connected to a special high-speed RoCE (RDMA over converged ethernet) network allowing for memory sharing on large-scale multi-GPU jobs, so users can also request multiple servers for analysis that requires more than four GPU cards.
The main objectives of this project are to examine the following questions:
- What is the distribution of per-GPU main memory usage over all runtime-weighed jobs in each partition? Knowing this will help ACCRE to understand our users memory needs for future hardware purchases.
- What is the distribution of the number of GPUs in each job (runtime-weighted) for each partition? What fraction of runtime-weighted and GPU-weighted jobs are using more than 4 GPUs and thus probably using the RoCE networking? Is this fraction different for each partition?
- What is the total runtime usage per-gpu (i.e. multiply runtime by the number of gpus) in each of the 3 partitions over the last year?
- What is the distribution of different groups and users accessing each partition? In each partition, who are the top users, and do they represent a majority of the runtime-weighted jobs on the partition?
- Currently there is a 5 day limit on runtime for GPU jobs, although some users have been asking for extensions. What is the distribution of requested runtime and actual runtime on jobs on each partition? Do users really need more time, or are they simply always requesting the maximum?