BITS Pilani, K. K. Birla Goa Campus, India
DaSH Lab (Data, Systems and HPC)
High Performance Computing (HPC)
DaSH Lab works on the data and file system aspect of HPC. Most emerging HPC applications have complicated I/O patterns which result in sub-optimal application performance.
Parallel File System
- How do you load balance I/O on to storage servers in a seamless manner?
- What is an optimal striping layout for a file within an application?
- Where to place files in a heterogeneous storage setup?
- How does network and I/O interact with each other in a parallel file system?
- How does I/O performance differ in parallel file system and object stores?
Collaborators: Virginia Tech - USA, Johannes Gutenberg University Mainz - Germany
BITS Pilani Personnel: Joel Tony, Yash Bhisikar, Bhavya Bajaj, Kaaviya Uthirapandian, Arnav Borkar, Hari Vamsi, Tushar Barman, Sreenath M.
Other Personnel: Debasmita Biswas, Sarah Neuwirth
PFS investigated: Lustre, IBM Spectrum Scale, BeeGFS, Ceph
Few Published Papers: ACM TOS 2024, Cluster 2023 (1), Cluster 2023 (2), INDIS@SC21, Cluster19, IPDPS19
Data Caching
- How to remove I/O bottleneck for large-scale deep learning applications?
- How can we use importance sampling for Distributed DL applications?
- Can I/O bottleneck be reduced for HPC applications by data prefetching?
Collaborators: Oak Ridge National Laboratory - USA, University of Viginia - USA, Virginia Tech - USA, Georgia State University
Other Personnel: Ahmad Maroof Karimi , Redwan Khan, Yue Cheng, Jong Choi, Lipeng Wan
Few Published Papers: FAST23, Cluster22
Resource Scheduling
- How can distributed deep learning jobs be made GPU heterogenity aware?
- How can distributed inferencing jobs be made GPU Heterogeneity aware?
- How can containers be used in HPC environments?
Collaborators: Virginia Tech - USA, Purdue University - USA, IIT Indore, Kuwait University, Northwestern Polytechnical University - China
BITS Pilani Personnel: Aditya Shiva Sharma, Kinshuk Goel, Amey Patil, Mukul Krishnan
Other Personnel: James Davis, Siddharth Sharma, Hadeel Albahar, Nannan Zhao
Few Published Papers: CCGrid22, Cloud20
Supercomputer Log Analysis
- How can darshan I/O characterization logs be used to characterize ML workloads?
- How can darshan logs be used for HPC trace generation?
- How are different supercomputing stacks used by varying HPC workloads?
- How do DL workloads differ in academia, industry and national labortories?
Collaborators: Virginia Tech - USA, Oak Ridge National Laboratory, Argonne National Laboratory, Lawrence Berkeley Laboratory
BITS Pilani Personnel: Natasha Meena Joseph, S Sai Vineet, Kunal Korgaonkar, Snehanshu Saha
Other Personnel: Ahmad Yazdani,Ahmad Maroof Karimi, Jong Choi, Phil Carns, Bing Xie, Suren Byna, Feiyi Wang, Jean Luca Bez
Few Published Papers: ICDCN23, MASCOTS21, HPDC22, HPDC22
Distributed Systems
DaSH Lab works on emerging areas of distributed systems, such as, serverless compute, blockchain, and cloud computing. All projects use state-of-the-art technologies like Kafka, Ethereum, Kubernetes, Docker.
Trust in Serverless
- How do you form a framework for FaaS using personal devices?
- How do you enable trust among the geo-distributed personal devices?
- How does one schedule FaaS on these devices?
- How is the cost of running FaaS provider services calculated?
BITS Pilani Personnel: Aalhad Sawane, Pranav Anand, Satyam Bansal, Pranay Varshney, Paras Mittal, Chinmay Rao, Amogh Balepur
Few Published Papers: Poster@HiPC23 (Best Undergraduate Poster Award)
Cross Cloud Compatibility
- How do you overcome a vendor locking situation with regards to public cloud providers?
- Can we have a middleware to have cross compatibility among major cloud providers, such as AWS, Azure and GCP?
Blockchain Smart Contracts
- How does one remove bugs in smart contracts after deployment?
- Are the upgrade mechanisms given by Ethereum usable for all smart contracts?
- Which upgrade patterns are most popular and why?
BITS Pilani Personnel: Aishwarya Parab, Nirmal Govindaraj, Siddhant Kulkarni, Pranay Varshney, Kunal Korgaonkar
Few Published Papers: Poster@HiPC23
Container deduplication
- What are the common layers in containers?
- Can common layers be deduplicated in large clusters using large amounts of containers?
- How can deduplication be solved in edge clusters with low resources?
- Can we devise caching methods for caching updates in edge clusters?
Collaborators: Virginia Tech - USA, Northwestern Polytechnical University - China, Kuwait University, IBM Research - USA, University of Minnesota - USA
BITS Pilani Personnel: Naman Agarwal, Prithvi Vishak, Nakul Bhachawat
Other Personnel: Hadeel Albahar, Nannan Zhao, Ali Anwar
Few Published Papers: ACM TOS24, TPDS20
Systems for ML
DaSH Lab tries to optimize systems and build frameworks for machine learning (ML) workloads. We aim at both large-scale ML workloads as well as tiny ML workloads suitable for resource-constrained edge devices.
Systems for Federated Learning (FL)
- How can multiple organizations collabate in a data-privacy manner in FL?
- Can knowledge distillation process be improved in FL?
- Are all edge devices needed to participate in a FL training process?
- Can we optimize energy in an FL process?
Collaborators: Queen Mary University of London - UK, Rochester Institute of Technology- USA
BITS Pilani Personnel: Sarang S, Druva Dhakshinamoorthy, Yuvraj Singh Bhadauria, Arihant Bansal, Pinki Yadav, Aditya Shiva Sharma, Sidharth C. Vivek, Vijay Dharmaji, Subroto Majumder, Manit Tanwar
Other Personnel: Ahmed Sayed, M. Mustafa Rafique
Few Published Papers: Poster@HiPC23
MLOps
- How to integrate automation of fault tolerance in MLOps pipeline?
- Can optimizations be done for deploying FL workloads using FLOps?
- Can intelligent Infrastructure-As-Code (IAC) be deployed for MLOps?
BITS Pilani Personnel: Aishwarya Jayashankar, Ayush Bharadwaj
Data Caching in TinyML
- How can distributed tinyML training be made aware of the heterogeneous edge environment?
- Can the datasets used for tinyML training be cached given the low resources on the devices?
BITS Pilani Personnel: Advik Raj, Sidharth C. Vivek
Data and I/O
DaSH Lab work on various data optimization techniques like compression or prefetching on various platforms like scientific workflows or database.
Data Compression
- Can we have a cost optimization model for running data compression algorithms on the cloud vs the end devices?
- How do the different optimization parameters of zstd fare against each other?
- How do large VM images fare with various levels of zstd?
Collaborators: Druva Inc.
BITS Pilani Personnel: Pinki Yadav, Vinayak Naik
DataSpaces
- DataSpaces is a data management framework for scientific workflows.
- Can optimizations be done for integrating HPC file systems into dataspaces?
- How can I/O intensive tasks be optimized in such a workflow?
Collaborators: University of Utah - USA
BITS Pilani Personnel: Joel Tony
Other Personnel: Manish Parashar
I/O Uring
- How can kernel I/O optimizations affect different large-scale use-cases?
- Which aspects of key value stores and databases can benefit from I/O uring optimizationsat the kernel level?
BITS Pilani Personnel: Arnav Gupta, Druva Dakshinamoorthy, Arjun Jagtap