Supercomputer Network R&D Engineer

Supercomputer Network R&D Engineer

Share

Full-time |AI Core System Development |DeepSeek|Hangzhou, Zhejiang

Apply for a position

Job description

What we need to do

You will participate in the design and implementation of cluster network architectures, covering cluster interconnection, routing, communication protocols, communication libraries, device firmware, and full-stack optimization for observability—this is a key factor determining the scalability of a ten-thousand-card cluster.

Specifically, it includes:

1. Cluster network architecture and protocol design

• Design ultra-large-scale cluster interconnection architecture and routing strategies to achieve packet-by-packet multipath load balancing and out-of-order packet processing mechanisms;

• Combined with on-site congestion control algorithms for business implementation, it addresses issues such as PFC storms and congestion spread.

2. Network observability

• Builds end-to-end monitoring and tracing systems (application layer metrics, latency, queues, packet loss, retransmission, etc.), leading network fault diagnosis and performance regression analysis for large-scale training.

3. End-to-end troubleshooting and performance optimization

• Familiar with computer architectures (CPU/GPU microarchitecture, device interconnection NVLink/UB/PCIe, memory models, etc.), and combine network loads for end-to-end profiling and full-link troubleshooting.

4. Cross-team collaboration

• Collaborate with teams in training frameworks, scheduling, storage, hardware, and other areas to drive continuous evolution of network performance.

We hope you have it

Basic background

• Bachelor's degree or above in computer science or related fields.

• Solid programming skills in C/C++/Python systems, familiar with network stacks and high-performance interconnect technologies (RoCEv2, IB, etc.).

Experience required

• Proficient in high-performance network topology, routing, and multipath load balancing design, understand out-of-order packet collection and reorder mechanisms, and master RDMA congestion control and related mechanisms.

• In-depth understanding of collective communication algorithms and principles, with experience in source-level development or deep performance tuning.

• A deep understanding of computer architecture allows you to clearly describe the entire process from memory to the end-to-end of a data packet, along with its potential bottlenecks.

• Possesses large-scale practical experience in network fault diagnosis and performance optimization, capable of locating issues across devices, switches, and network cards.

Bonus items

• Experience designing or long-term operation of GPU cluster networks with over 10,000 cards.

• Experience in high-performance computing HPC application development, having run large-scale parallel tasks at the full machine level.

• Published papers on high-performance networks or distributed networks at top conferences in systems such as SIGCOMM, NSDI, SC, ISCA, and OSDI.

• Experience in firmware/OS development for smart network cards and programmable switches (P4, SDN, SONiC, etc.), familiar with hardware offloading and acceleration pipeline design.

• Familiar with the development of heterogeneous acceleration devices such as DPU, IPU, and FPGA.

You will get it

• Directly defining and implementing next-generation supercomputing network architectures, challenging the ultimate performance of larger-scale clusters.

• Work with top system/operator/framework engineers.

• Your work will directly affect the efficiency and cost of large model training and inference.

Job information

Job title

Supercomputer Network R&D Engineer

Salary range

-

Nature of the position

Full-time

Types of functions

AI core system development

Department

DeepSeek

Work location

Hangzhou, Zhejiang