Portable Acceleration of CMS Computing Workflows with Coprocessors as a Service

Area of Science:

• High-energy physics computing
• Machine learning infrastructure

Background:

• Scientific experiments like the CMS experiment at CERN's LHC face escalating computing demands.
• Central processing units (CPUs) alone may not meet future performance needs.
• Coprocessors offer architectural advantages for specific computational tasks.

Purpose of the Study:

• To explore the Services for Optimized Network Inference on Coprocessors (SONIC) approach for large-scale data processing.
• To evaluate the deployment of SONIC as a service for scientific workflows.
• To demonstrate the benefits of offloading machine learning inference tasks to coprocessors.

Main Methods:

• Implemented a data processing workflow from the CMS experiment.
• Executed the main workflow on CPUs and offloaded machine learning inference tasks to Graphics Processing Units (GPUs) as coprocessors.
• Conducted experiments across Google Cloud and the Purdue Tier-2 computing center.

Main Results:

• Achieved individual acceleration of machine learning algorithms on coprocessors.
• Demonstrated significant throughput improvement for the entire data processing workflow.
• Validated the SONIC approach's effectiveness in diverse cloud and local computing environments.

Conclusions:

• The SONIC approach effectively accelerates machine learning inference in scientific data processing.
• SONIC enables high coprocessor utilization and workflow portability across different coprocessor types.
• This method can be generalized to various coprocessors and deployed on local CPUs without performance degradation.