Worked on Distributed Storage System

Part of GPU team for AHV Hypervisor1st place Winner at Nutanix U-Hack Intern Hackathon◦ GPU Virtualization Driver Automation: Spearheaded the automation of integrating newly released NVIDIA vGPU Drivers, advancing GPU Virtualization in Nutanix's AHV Hypervisor. This automation significantly decreased the Time-To-Market for new drivers, reducing it from 72 hours to just 8 hours, making it available to customers 9 times faster◦ GPU Virtual Functions in L1 RHEL Hypervisor: Conducted a proof of concept (POC) for creating Virtual Functions (VFs) of GPUs (i.e., vGPUs) by utilizing their SR-IOV capabilities within an RHEL L1 Hypervisor, which operated as a guest under an AHV Hypervisor. The GPU on the AHV Host was passed through to the RHEL Guest Hypervisor (L1) to ultimately provide vGPUs to L2 Guest VMs◦ Enhanced Bug Reporting: Automated the installation of the RHEL 9.2 Hypervisor on bare-metal Nutanix hardware and configured the KVM reference on the hypervisor. This setup was used to reproduce Nvidia bugs and enabled prompt bug reporting◦ Improved Code Coverage of CI: Improved the code coverage of the AHV Hypervisor by identifying untested code areas and incorporating them into the test suite, which is part of the continuous integration pipeline for development

◦ Add NVMe device support to storage stack: I was part of a team that was responsible for making performance changes to the storage stack to make the code Lock-less and leverage NVMe SPDK framework’s Direct Memory Access capabilities by using huge pages for memory buffers to avoid memory copying from kernel-space to user-space processes and vice-versa. Lock-less implementation was achieved by using ring buffers for message-passing among multiple threads for synchronization◦ I/O performance optimization: Part of a team that was responsible for making performance optimizations to the Robin CNP I/O pipeline. The optimizations involved batching write I/Os to be written to the same device and coalescing write I/Os according to the offsets on the disks.◦ Persist Garbage Collector Configs: The task involved persisting garbage collector configs on PostgreSQL DB, to avoid GC configs being initialized to their defaults on system restarts◦ Rolling upgrade of the platform: Led the efforts to enable the rolling upgrade of the storage stack on a cluster. This involved handling communication among inconsistent RPC versions on storage and compute nodes.

◦ Took the lead in creating REST API server in Node.js(Express framework) for famstar's brand portal and influencer flutter application◦ Designed the system end-to-end from scratch, which involved gathering requirements, documenting them, and designing DB schemas◦ Created transactional queries to maintain atomicity in several sensitive DB operations◦ Created MongoDB Aggregation Pipeline queries to serve millions of records efficiently

◦ Authorization: Developed a C/C++ library supporting Access Control List(ACL), Role Based Access Control(RBAC) and Attribute Based Access Control(ABAC) authorization paradigms◦ Batch Persistent Storage: Implemented batch addition and removal of policy rules and integrated with various persistent storage adapters◦ Middlewares: Developed authorization middlewares for various HTTP REST API frameworks like Express and Koa

◦ Prominent Key-Phrase Extraction: Devised a Bi-RNN(Bidirectional Recurrent Neural Network) model for extracting prominent key-phrases from unstructured textual data of social-media micro-blog text domain◦ On-Device Optimizations: The model was optimized to execute on a limited resource device to avoid server computations and enable On-Device computation and extraction of key-phrases◦ Unsupervised Models: Experimented with unsupervised-learning models involving algorithms like TopicRank and TextRank, with POS(Part-of-Speech) tagged tokenization using regular expressions