Exploiting GPUs in virtual machine for BioCloud.

Area of Science:

• Computational Biology
• Cloud Computing
• High-Performance Computing

Background:

• Biological applications increasingly utilize Graphics Processing Units (GPUs) for enhanced computation.
• Cloud environments offer scalable resources but often lack direct GPU access for virtual machines (VMs).
• Existing GPU sharing mechanisms in VMs do not meet the throughput demands of biological applications.

Purpose of the Study:

• To propose a BioCloud system architecture enabling VMs to directly utilize GPUs in a cloud environment.
• To achieve near-native performance for biological applications by bypassing traditional sharing methods.
• To implement an efficient GPU multiplexing strategy for VMs.

Main Methods:

• Exploited the PCI Express (PCI-E) pass-through mode for direct VM-to-GPU access.
• Implemented GPU multiplexing using PCI-E's hot plug-in/out device features for on-demand GPU allocation.
• Utilized Xen VMM and NVIDIA GPUs for system implementation and testing.

Main Results:

• The BioCloud system allows VMs direct access to underlying GPUs, achieving performance comparable to native environments.
• On-demand GPU time-sharing is enabled through hot plug-in/out, optimizing resource utilization.
• The prototype demonstrated high effectiveness for biological GPU applications in the cloud.

Conclusions:

• The proposed BioCloud architecture successfully enables efficient and high-performance GPU utilization for biological applications in cloud computing.
• Direct GPU access via PCI-E pass-through is crucial for meeting the computational demands of these applications.
• The system offers a viable solution for researchers to leverage cloud resources for intensive biological computations.