Jamie Liu Email and Phone Number

1. Kubernetes AI platform for offline trainingRun ML frameworks like TensorFlow, PyTorch, pySpark with jupyter notebook in docker container, retrieve training data from S3 object storage, get CSV file from Hadoop platform, manage training task with kubernetes job. Support Spark on kubernetes, horovod distributed training framework and volcano job scheduling.2. Kubernetes AI platform for online inferenceUse S3 object storage as model repository, deploy AI models with tfserving and torchserve model server, support multi-model serving, dynamic model loading and unloading and live model update.Build a common model inference platform with kfserving, serving ML frameworks like Tensorflow, PyTorch, PMML and ONNX. Support istio traffic management, blue-green deployment and canary release.3. GPU management and virtualizationManage GPU as a resource pool within kubernetes cluster, and allocate GPU resource according to the GPU utilization. Hook CUDA driver API to isolate device memory and switch CUDA context for GPU virtualization. Share GPU among all nodes of the cluster by rCUDA (Remote CUDA).4. Blockchain BAAS platform based on KubernetesBuilt a Hyperledger fabric blockchain verification platform, and developed a blockchain service gateway for channel and chaincode operations. It was used for immutable ledger storage scenarios such as asset certification and labor wage payment.

1. Ceph object storage for massive small files and Persistent Volume provisioningTune Ceph configuration to improve performance, cosbench stress test showed the IOPS was increased by nearly 2 times. It had stably carried more than one billion small pictures, and the daily increment was nearly 5 million. RBD random write 4k block IOPS increased by nearly 40%. Located the cephfs kernel client performance issue during synchronous IO. Dynamic volume provisioning for kubernetes allowed ceph RBD and glusterfs volumes to be created on-demand.2. Real-time log analysis with pySpark streamingCollected container json-file logs with fluentd and pushed to kafka for consumption. Logs were pushed into elasticsearch for query, and analyzed by spark streaming to generate statistics such as QPS and latency, which written into redis for real-time alarm.

1. Open vStorage converged storage for Virtual MachinesOpen vStorage, similar to VMware vSAN, was a grid storage router. It turned Swift object into block storage, SSD cache was used to ensure the read and write performance. As the key contributor, responsible for porting the code to Centos 7 and supporting systemd manager. Tuned performance and found out the bottleneck during md5 hash dedup, avoided the issue by taking dedup for os image volume and non-dedup for data volume.2. High-performance NAS storage system based on glusterfs and ganeshaGlusterfs as the backend storage and ganesha as the NAS server, built high-availability NAS service through LVS and keepalived load balance. Implemented Rest API management interface with python flask. Solved performance issue, located NFS client hang problem caused by LVS TCP long connection.

1. Storage server integration testSystem Verification: Define test requirement, verify the integration of server modules such as BMC, BIOS, POST, etc. Run automation testing to verify storage platform performance and reliability, analyze System Event Log, locate system hardware and firmware issues. Access backend storage array via iSCSI HBA and run end-to-end test, measure the IOPS and throughput of the VNX storage platform.2. Develop server storage diagnostic systemNetwork Diagnostic: review HW design spec and release plan, define test requirement, write user space device driver for traffic test, solve the problem of random packet loss on the Management Module Control Stations.Secondary development of Vendor API: write Vendor API specification, verify and integrate vendor API library into EMC diagnostic system. Locate issues found during integration and performance testing, and work closely with Vendor to reproduce and resolve issues.ODM Project: implement traffic test based on linux socket, performance test with iperf, debug with wireshark.

1. Ethernet Driver Development: Based on P4080 communications processors with data path, configure Data Path Acceleration Architecture HW, write 1G and 10G Ethernet driver, support 8 OS clients to handle traffic independently by hypervisor mechanism, and work as a virtual Switch.2. Multi-site teamwork: team members are from USA, Sweden and Australia，run the project in agile model, which deeply improves the communication and team work skills.

1. Desktop Daemon: Handle D-Bus method call request from other processes，monitor udev system events, broadcast event to registered applications when system configuration changes.2. Applet Framework Design: Make applet as a Dynamic Link Library, design the applet framework based on MVC design pattern, implement system configuration applet.