Big Data application scheduling and load balancing
Heterogeneous computing platforms that take advantage of CPUs, GPUs, and energy efficient FPGAs continue to be adopted as computational needs increase across applications in finance, deep learning & AI, genomics, and machine vision. Integration and management of these platforms for big-data computing frameworks like Apache Spark™ can become a complex task that requires synchronization between Big Data application developers, accelerator designers and system administrators.
Delivered as a two-component to eliminate this challenge, Falcon Computing’s Kestrel™ Runtime is an accelerator management tool for heterogeneous compute clusters. It provides Apache Spark™ compatible accelerator virtualization, monitoring, queuing, allocation and scheduling.
Kestrel Runtime Performance
The Global Accelerator Manager (GAM) is built upon Apache YARN and provides global resource allocation within a cluster so as to optimize system throughput and accelerator utilization.
The Node Accelerator Manager (NAM) is built upon Blaze runtime system, which is an open-source project from UCLA that allows Spark applications performance improvement with hardware accelerators. It provides an accelerator virtualization layer for applications and handles accelerator sharing, isolation, and accounting.
Built for ease-of-use
- Container-based accelerator reservation
- Virtualization with automatic accelerator allocation and scheduling
- Simple APIs to invoke accelerators in applications
Optimized for performance
- Fine-grained accelerator sharing
- Data caching and data-transfer/computation overlapping
- Accelerated Machine Learning Lib (MLlib)
Designed with scalability in-mind
- Support heterogeneous systems including FPGA and GPU
- Easy integration with application frameworks
- Built-in integration with Apache ecosystem