Chijun Sima

Ekko: low-latency model update for multi-terabyte DLRMs (published in part as OSDI '22)

• Problem. Scaling DLRMs improved offline accuracy but degraded online engagement; root cause: stale models from increased model-update latency.
• Key idea. Co-designed deployment mechanisms with model-aware policies (compressed update dissemination, accuracy-aware scheduling, SLO-aware placement, safe rollback).
• Technical contributions. WAN bandwidth −92 %, machine cost −49 %, 2.4 s model-update latency; 10,000× model-size scaling (GB → tens of TB).
• Outcomes. Core techniques published as OSDI '22 (co-first author). Deployed in WeChat recommendation stacks, serves 1 B+ users daily. Official WeChat blog reports +40 % DAU and +87 % total VV over six months after full adoption (alongside product iteration and operations).

Data and feature platform: safe, scalable pipelines

• Problem. Modern feature pipelines are long and increasingly multimodal; cross-process operator composition creates high overhead and expensive data movement.
• Approach. WebAssembly-based runtime for in-process isolation (safety + resource constraints) and locality-aware operator placement near data sources.
• Outcome. Data movement reduced up to 1,200× on representative workloads; widely used within WeChat for data preparation.

LLM serving systems

• Building cost-effective serving mechanisms around remote KV-cache storage and compression.

• Improved Semi-NCA performance and optimization pipeline; shipped in LLVM 9.0 (reported speedups up to 1,980× on real-world samples).
• Unified APIs on dominator trees; shipped in LLVM 7.0.