Tech Lead and Manager for Cross-Network Communication products in the Google Compute Engine’s Networking Control Plane. As tech lead, own the product definition (create the resource model that users interact with, and write user guides jointly with product management), architecture, high-level design, component design reviews, and code reviews. As individual contribute component design and implementation for some components of my products. As manager own the engineering roadmap for the products where multiple teams delivers the components of the high-level design. The teams do the component-level design and implements the REST API that exposes the resource model, validates the user requests to guarantee correctness, and translates the user requests to data plane configs. The organizing principle of the work is to allow consumers to access producer services (Google-managed or Third-Party running on GCP, or Google-managed running on GCP or Borg) using private IP address so that consumer machines don’t need a public IP address and traffic stays internal to Google cloud. The systems provide two sets of APIs: for end consumer, and for service producers.Service Connectivity Automation – Provides easy access to Google-managed services (e.g., Cloud Redis) which are built on Private Service Connect (PSC). Automatically configure (create resources and enforce policies on) consumer networks for the GCP service producers (e.g., Cloud SQL, Cloud Memory Store). PSC Propagation – For consumers of Network Connectivity Center, when one spoke connects to a Google-managed service, easily allow other spokes in the hub to connect to the service too.Peering to PSC Migration – Automated cutover of traffic from VPC Peering to PSC without changing consumer clients; requires transferring IP addresses from over VPC to its peer.Terraform Automation – Set up all the resources needed to connect to different managed services either from a VPC network or from on-prem.

Private Service Connect Endpoints – Built on Network Address Translation where the NAT is transparent to the consumer. Producer exposes a service. Consumer specifies an IP address on their network and the service it wants to talk to, and all traffic to the address shows up at the producer. Private Service Connect Endpoints – Producer exposes a service. Consumer specifies an IP address on their network and the service it wants to talk to. There are multiple products that allow consumers to access Google services either on GCP (Cloud SQL, Apigee, Vertex) or Borg (Big Query), or third-party producers (MongoDB), that use either L4 or L7 Load Balancing.Private Service Connect Interfaces – Allows a VPC to give an IP for use on a virtual machine in another org or project. This allows producers virtual machines to have a network interface which grants the producer to have access to the consumer VPC.Private Service Access – Provides customers easy access to Google-managed services (e.g., Cloud SQL) built on Network Peering. Abstracts away the use of Network Peering between consumer network and a Google owned network, handles the IP range management, and manages the shared network on Google side on behalf of multiple producer services.

Delivering features that help Oracle compete in the IaaS public cloud landscape as part of the Control Plane Team for Virtual Networking for Next Gen Oracle Cloud.Feature 1: Internal DNS - It allows compute instances in a customer's virtual network to communicate with each other using host names, instead of having to rely on either using explicit IP addresses, or having to manage host files. Feature 2: Network interface routes for external destinations - These allow all traffic leaving a customer's virtual network for the outside to go through a compute instance in the virtual network. This enables the virtual network user to run a network appliances that can inspect the traffic entering and leaving their network.

Delivering architecture and engineering to obtain quantified measures of IaaS cloud quality. The strategy drives on two initiatives: 1) • Health-Model: deliver metrics (TopN, Hotspots, Outliers) to optimize cloud performance - pipeline to collect, store, analyze, and present data. 2) • Platform Skew: ensure consistency of software across multiple clouds - system to allow actions (install, upgrade, modify), log results (success, failure), show state (versions, history)Directed complex performance analysis work in diverse scenarios.1) Virtual Machine Disk I/O: factors affecting the VM IOPS on KVM hypervisor (filesystem - ext3 vs. ext4, image - raw vs. qcow2, VirtIO – cache vs. writeback, native block access).2) Virtual Machine Launch: get time spent in various constituent parts of the boot process (Keystone, Nova, Glance, and Neutron) by instrumenting OpenStack Rally.Programming Models (Technologies): IaaS (OpenStack), Virtualization (KVM), Containers (LXC), Automated Deployment (Ansible), Data Analysis (Python Pandas, Kibana), NoSQL Databases (MongoDB, ElasticSearch), Testing (PyTest), Continuous Integration (Git), Continuous Deployment (Jenkins)

Developed software for Intel Education Kno platform (https://www.kno.com/) as member of cloud engineering team.Demonstrated an ability to quickly understand an existing large distributed system, and to contribute with quality both, new features and improvements to existing features. This required meeting the challenge of figuring out the system at multiple levels: system architecture at the high-level, the relevant parts necessary for my contributions at component-level, and finally the existing code base written by other people at development-level. Projects1) Student Roll Synchronization: Created a library that extracts student roll for sections from two systems, does tree traversal for school district hierarchies, and performs set arithmetic on the section student data to determine necessary updates. Solution interfaces with other REST API, messaging queue, and database systems. 2) User Content Deletion: Delivered an REST API that allows students and teachers to delete documents they upload to the system. It updates the user account information relevant to the deleted document, and updates the state for other devices on which the user has the same content. 3) Performance Tuning: Optimized the book purchase workflow. Crafted hypothesis, created client driver, instrumented server-side code, and tested changes. Improved message queue publish times, reduced the number of database queries, and determined missing database indexes.Programming Models (Technologies) – RESTful Services (Java, JSON, Jetty), Databases (MySQL), Publish/Subscribe (Rabbit MQ) , Testing (JUnit) , Continuous Integration (Git).

Lead the Cloud Analytics team for Intel Private Cloud. Owned the product architecture, and engineering of pipelines with millisecond data delivery guarantees. Contributed to the team as a software engineer.Main challenge was to provide a enterprise-wide business-relevant view of the data assembled from details about independent systems that provide compute and storage resources, and are spread across sites and data centers. Worked on a performance analysis, capacity management, and centralized logging solutions for Openstack cloud based on KVM hypervisors and Ceph storage servers. Performance data pipeline: Collectd collector, MongoDB aggregator, and Graphite visualizer. Capacity data pipeline: custom scripts to query Openstack API for inventory data, correlate it with performance data, aggregate the results in MongoDB, and provided output using REST API and a web front-end. Log aggregation: Syslog, Logstash, ElasticSearch and Kibana. Developed alternative solution using Syslog, Flume, and Hadoop for comparison. The ELK version won out due to ease of use.Demonstrated an ability to comprehend and debug complex distributed systems by making code changes to resolve data collections issues in three releases of OpenStack Ceilometer.1) Folsom – reconciled disconnect with virsh domain model for network interfaces.2) Grizzly – modified the code to determine VMs running on a hypervisor.3) Havana – corrected the call to connect to Glance over HTTPS.Demonstrated analytics expertise with OpenStack compute host load analysis. Developed a tool to create Nova virtual machines with Cinder volumes and Nova networks on target host, and correlated host behavior with data from system activity reporter (sar) that is part of sysstat.Programming Models (languages): IaaS (OpenStack), NoSQL (MongoDB, ElasticSearch).

Led a team of database architects and engineers that delivered a capacity and performance data solution for an internal custom cloud based on VmWare Virtual Centers and EMC Storage Area Networks. The biggest challenge we faced here was bridging the semantic gap between the data in different sources.The solution used classical SQL database ETL approach. Data was extracted from the backend databases for the virtualization and storage systems, transformed, and then loaded to the user-accessible database. The transformation involved joins and aggregations. User access was through Microsoft Excel files with embedded queries.Programming Models (Technologies): SQL (Microsoft), Virtualization (VMWare), SAN (EMC).

Focused on developing manageability solutions to improve operational efficieny of data centers supporting Intel 300mm factories. 1) Member of team that delivered a 25% reduction in operations head count for data center. Owned the analysis of Windows, Storage, and Backup and Restore tasks, and solution recommendations. Reduced MTTR by delivering monitoring solutions using Microsoft System Center Operations Manager 2007 for 75 factory automation applications that run in the data center. Developed libraries for file, folder, and cluster monitoring to compensate for shortcomings in default Microsoft solutions. Improved organizational strength by training ~50 people to develop monitoring solutions over 3 years. 2) Participated in team to consolidate number of servers that run factory applications. Responsible for 19 application consolidations that drove their server count from 50 to 19. Co-engineered introduction of Microsoft Cluster Server, and solely-engineered Microsoft x64 OS solution for application consolidation. 3) Owned the decision on hardware landing of new factory automation applications, application performance reviews, and application interaction with hardware as member of team that owned architecture standards for all data center applications. 4) On operational front led a team that decommissioned 350 servers in the TD data center without impact to development factory. Led to reduction in maintenance and power costs, and space savings in data center.

Conceptualized, designed, analyzed, implemented and evaluated system support for programming distributed computing applications. The projects I have worked on are:1)Reasoning with space, time and identity in distributed applications: System support for applications in diverse domains (e.g. surveillance, traffic management, mobile commerce) that need this reasoning ability. An example from surveillance is correlating a person with a location to determine if the times he has been there cause suspicion. The system also involves node communication, participation protocol for application nodes, and data storage for reasoning about the past.2)Programming idioms for pervasive computing: System support to handle temporal stream data across distributed, heterogeneous nodes – multiple languages (C, Java) and operating systems (Linux, Windows 2000). The system is fault tolerant as it can handle node failures, and notify surviving nodes about the failures. It supports dynamic node participation by allowing a node to join and leave without affecting other nodes. 3)Performance study of a cluster runtime system: Instrumented a runtime system to account for application performance at the granularity of processor clock cycles. The total time was split into time spent in application logic, runtime system, and communicating across the network.

Designed and implemented an IP-blocking module for a file-sharing system at open-source site Sourceforge. This involved developing the front-end for input of blocking information, the actual blocking module, and integrating these two pieces with the existing system.

Designed and implemented a web-based information management system for a computer grid that provides various mechanisms -- volunteer computers for use in grid, search for suitable computers, reserve the computers, remove the computers from the grid.

Designed and implemented a prototype for third-party interactions, using Websphere e-commerce system. This involved understanding the existing two-party system (e-commerce protocols, system architecture and implementation), and then creating a compatible three-party system.

Maintained and added features to an analysis tool for a network of frame relay switches.Designed and implemented a communication layer, of distributed-system testing tool.