Saman Biookaghazadeh

About Saman Biookaghazadeh

At Google Research, I work on building the next generation of language models, more specifically diffusion language models.Here\'s the summary of my past works:20••••25: I contributed to the development of Google Cloud\'s first A3+ GPU cluster, utilizing H100 GPUs for large-scale training workloads up to GPUs. This project included the development of a proprietary RDMA technology to facilitate high-performance GPU-to-GPU communication within the NCCL library.20••••25: I developed a Google-specific GPUDirect solution for direct storage-to-H100 GPU data loading, integrated with tf.data and now used by Google DeepMind.20••••24: I led a cross-functional project to optimize Google\'s storage infrastructure, reducing both data volume and latency. Our \"weighted chunk allocation\" algorithm, which balances I/O across drives with varying performance, has saved Google millions of dollars.Before entering industry, I pursued my Ph.D. in High Performance Computing, where my research centered on optimizing distributed computation and storage systems. My current academic work explores scheduling distributed workloads across heterogeneous systems (CPUs, GPUs, and FPGAs) to enhance performance, with a specific focus on the demanding computational needs of Big Data and AI.

Experience

1. Research Engineer

   Google DeepMind

   Oct 2023 — Present · NY, US

Education

• Arizona State University

  Doctor of Philosophy (PhD), Computer Systems

  2015

• Florida International University

  Doctor of Philosophy - PhD, Computer Science

  2013 — 2015

• University of Tehran

  Bs, computer engineering

  2009 — 2013

Skills

• J2ee Application Development
• C
• Spring
• Web Development
• C++
• Linux Kernel Development
• Xen and Kvm Hypervisor
• Python
• Linux
• Algorithms
• Computer Science
• Java