Tensor Cores architecture can go beyond AI/ML! Matrix multiplication has many other siblings, including dynamic programming algorithms and minimum spanning tree problems that share the same computation pattern but different arithmetic operations! Andrew Zhang presented our work at ISCA 2022, called SIMD^2 that uses NVIDIA’s Tensor Cores to create a set of 2-dimensional SIMD instructions in supporting these matrix semiring problems. We achieved 10x speedup over state-of-the-art SIMD/CUDA implementations with just 5% area overhead!
Very excited to share with you our latest open-source project on using Google’s Edge TPUs for general-purpose computing! Our project releases the power of matrix processing on Edge TPUs to any application/problem that you could use matrix algebras to implement/solve (as long as you can tolerate a little bit of inaccuracy due to Edge TPU’s limited precision).
You can now clone our project from the following GitHub repo
Enjoy and let us know what matrix application/matrix you can come up with!
NDS @ Non-volatile memory workshop
The ESCAL group attended their first “in-person” event since March, 2020. The long-awaited non-volatile memory workshop @ UCSD. Yu-Chia presented his best paper-nominee work, NDS, during the best paper session (again!) at NVMW. If you’re interested at NDS, you may find our paper at https://dl.acm.org/doi/10.1145/3466752.3480122
and our code release at https://github.com/escalab/NDS
TCUDB @ SIGMOD
Tensor cores can do more than AI/ML! Yu-Ching Hu presented our paper “TCUDB: Accelerating Database with Tensor Processors”, at SIGMOD 2022! TCUDB is a database engine using Tensor Cores on NVIDIA’s latest Ampere architecture to accelerate entity matching, graph query processing, and matrix-based data analytics by up to 288x compared against conventional GPU database systems.
With the rapid growth of dataset sizes but limited improvement of high-performance computers, we need to revisit the existing programming and execution models to efficiently utilize all system components. In modern computers, lots of deficiencies in applications are related to data management and movements. The vision of Extreme Storage & Computer Architecture Laboratory is to revolutionary change the way how people think about programming and computing today — using a data-centric perspective in programming instead of the conventional computing-centric approach. ESCAL conducts research in systems and computer architecture with focus on storage systems, parallel processing, high-performance computing, programming languages and runtime systems.
Research Projects
Accelerating non-AI/ML applications using AI/ML accelerators
The explosive demand on AI/ML workloads drive the emergence of AI/ML accelerators, including commercialized NVIDIA Tensor Cores and Google TPUs. These AI/ML accelerators are essentially matrix processors and are theoretically helpful to any application with matrix operations. This project bridges the missing system/architecture/programming language support in democratizing AI/ML accelerators. As matrix operations are conventionally inefficient, this project also revises the core algorithm in compute kernels to better utilize operators of AI/ML accelerators. With this project, ESCAL envisions ourselves to lead the next trend of a revolution — similar to the one happened on GPUs. You may now try our most recent GPTPU project from the GitHub repo: https://github.com/escalab/GPTPU
As parallel computer architectures significantly shrinking the execution time in compute kernels, the performance bottlenecks of applications shift to the rest of part of execution, including data movement, object deserialization/serialization as well as other software overheads in managing data storage. To address this new bottleneck, the best approach is to not move data and endow storage devices with new roles. Morpheus is one of the very first research project that implements this concept in real systems. We utilize existing, commercially available hardware components to build the Morpheus-SSD. The Morpheus model not only speeds up a set of heterogeneous computing applications by 1.32x, but also allows these applications to better utilize emerging data transfer methods that can send data directly to the GPU via peer-to-peer to further achieve 1.39x speedup. Summarizer further provides mechanisms to dynamically adjust the workload between the host and intelligent SSDs, making more efficient use of all computing units in a system and boost the performance of big data analytics. This line of research also helps ESCAL receive Facebook research award, 2018 and MICRO TopPicks in 2020.
Efficient storage system for heterogeneous servers
Although high-performance, non-volatile memory technologies and network devices significantly improve the speed of supplying data to heterogeneous computing units, the performance of these devices are still far behind the capabilities of heterogeneous computing units. For example, modern SSDs can read more than 3GB of data per second, but GPUs can process more than 17GB of data for database aggregation operations within the same period of time. As result, the heterogeneous computing units are under-utilized. We will revisit the design of existing runtime systems to transparently improve the utilization of system components, potentially leading to speedup or better energy-efficiency.
Optimizing the I/O system software stack for emerging applications
With hardware accelerators improving the latency in computation, the system software stack that were traditionally underrated in designing applications becomes more critical. In ESCAL, we focus on those underrated bottlenecks to achieve significant performance improvement without using new hardware. The most recent example is the OpenUVR system, where we eliminate unnecessary memory copies and allow the VR system loop to complete within 14 ms latency with just modern desktop PC, existing WiFi network links, raspberry Pi 4b+ and an HDMI compatible head mount display.
Yu-Chia Liu and Hung-Wei Tseng, NDS: N-Dimensional Storage. In the 54th Annual IEEE/ACM International Symposium on Microarchitecture, MICRO 2021. [GitHub] [Best Paper Nominee]
Yu-Ching Hu, Murtuza Lokhandwala, Te I and Hung-Wei Tseng. Dynamic Multi-Resolution Data Storage. In the 52nd Annual IEEE/ACM International Symposium on Microarchitecture, MICRO 2019. [Best Paper Honorable Mention]
Kiran Kumar Matam, Gunjae Koo, Haipeng Zha, Hung-Wei Tseng and Murali Anavarum. GraphSSD: Graph Semantics Aware SSD. In the 46th International Symposium on Computer Architecture, ISCA 2019.
Gunjae Koo, Kiran Kumar Matam, Te I, Hema Venkata Krishna Giri Narra, Jing Li, Steven Swanson, Murali Annavaram, and Hung-Wei Tseng. Summarizer: Trading Bandwidth with Computing Near Storage. In 50th Annual IEEE/ACM International Symposium on Microarchitecture, MICRO 2017
Yanqin Jin, Hung-Wei Tseng, Steven Swanson and Yannis Papakonstantinou. KAML: A Flexible, High-Performance Key-Value SSD. In 23rd International Symposium on High Performance Computer Architecture (HPCA 2017). February 2017.
Yang Liu, Hung-Wei Tseng, Mark Gahagan, Jing Li, Yanqin Jin and Steven Swanson. Hippogriff: Efficiently Moving Data in Heterogeneous Computing Systems. In 34th IEEE International Conference on Computer Design (ICCD 2016). Oct. 2016.
Yang Liu, Hung-Wei Tseng and Steven Swanson. SPMario: Scale Up MapReduce with I/O-Oriented Scheduling for the GPU. In 34th IEEE International Conference on Computer Design (ICCD 2016). Oct. 2016.
