What is Mainstream?

Mainstream is a video analysis system that jointly adapts concurrent applications sharing fixed edge resources to maximize aggregate application result quality.

Paper: https://www.usenix.org/system/files/conference/atc18/atc18-jiang.pdf

Why Mainstream?

Many practitioners use transfer learning and fine-tune a pre-trained model for their dataset instead of training their DNN (deep neural network) from scratch. Since the same few common pre-trained models are used, we can share computation between concurrent applications that operate on the same video stream.

Since the decision of how much to share is dependent on the co-located applications and the hardware of the edge device, it is important that the scheduling decision be made at deployment time rather than by the application developer at training time. Mainstream does this automatically to get the best aggregate performance across applications.


Mainstream Architecture

M-Trainer takes a labeled dataset and outputs an M-Package. M-Scheduler takes independently generated M-Packages, and chooses the task-specific degree of specialization and frame rate. M-Scheduler deploys the unified multi-task model to M-Runner, performing inference on the edge.


CMU Faculty

Intel Research Pittsburgh

PhD Students

Undergraduate Students

CMU students and academics: join us!

If you’re at CMU, we would love for you to work with us!

Systems for ML. Scheduling. Multi-Tenant Video Processing – at the edge. Distributed systems. Fairness. If this strikes a chord with you, let us know!

For students, whether you’re an undergraduate or graduate student, looking for an independent study project, a MCDS project, a capstone project, a class project, or a summer research attachment, feel free to link up with us.

Contact Us

If you have feedback or am interested in collaborating, feel free to reach out to us.


Mainstream: Dynamic Stem-Sharing for Multi-Tenant Video Processing.
Angela H. Jiang, Daniel Lin-Kit Wong, Christopher Canel, Lilia Tang, Ishan Misra, Michael Kaminsky, Michael A. Kozuch, Padmanabhan Pillai, David G. Andersen, Gregory R. Ganger
USENIX ATC 2018. [PDF] [BibTeX]

Dynamic Stem-Sharing for Multi-Tenant Video Processing.
Angela Jiang, Christopher Canel, Daniel Lin-Kit Wong, Michael Kaminsky, Michael A. Kozuch, Padmanabhan Pillai, David G. Andersen, Gregory R. Ganger
SysML Conference 2018. [PDF]


We thank the member companies of the PDL Consortium (Alibaba, Broadcom, Dell EMC, Facebook, Google, HPE, Hitachi, IBM Research, Intel, Microsoft, MongoDB, NetApp, Oracle, Salesforce, Samsung, Seagate, Toshiba, Two Sigma, Veritas and Western Digital) for their interest, insights, feedback, and support. This work is supported in part by funding from Intel as part of the Intel STC for Visual Cloud Systems (ISTC-VCS).