Blaze: Deploying Accelerators at Datacenter Scale

Description

Blaze is an accelerator-aware programming framework for warehouse-scale accelerator deployment. Blaze provides a programming interface that is compatible to Spark, an in-memory compute engine for large-scale data processing, and a runtime system that provides transparent accelerator management. With Blaze, the deployment effort of accelerator task is reduced by more than 10x compared to traditional frameworks such as OpenCL. Blaze is designed to efficiently manage accelerator platforms with intelligent data caching and task pipelining schemes to minimize the communication and data movement overheads.

License

The Blaze runtime system is released under BSD 2-Clause license. The Blaze runtime system includes the following components:

• AccRDD: A Spark extension including interface between Spark program and accelerators
• manager: A accelerator manager that controls all local accelerators including context setup, data transfer and task execution

In addition, each accelerator task is compiled with Blaze APIs. Some simple examples are provided in the ./examples folder, which are released without restriction.

Installing Blaze

0. Prerequisites 0. Boost (tested with 1.55.0) 0. Google Protobuf (tested with 2.5.0) 0. Apache Spark (tested with 1.3.1) 0. Apache Hadoop (tested with 2.5.2)
1. Compiling 0. Edit Makefile.config 0. export BLAZE_HOME=$PWD/accrdd 0. run make

Running Loopback example

0. Prerequisites 0. Intel AALSDK 4.1.7 0. Intel AAL NLB example RTL
1. Compiling 0. From root directory run make examples
2. Execution 0. Start NLB ASE 0. On terminal 1 start Manager: ./manager/bin/acc_manager ./examples/loopback/acc_task/conf.prototxt 0. On terminal 2 tart Spark program: ./examples/loopback/app/run.sh or ./examples/loopback/app/run_local.sh

Contacts

For any question or discussion, please contact the authors:

• Di Wu: [email protected]
• Hao Yu (Cody): [email protected]