Empowering systems design with end-to-end specialization

Tuesday, October 3, 2017 -
1:00pm to 2:00pm
1240 CS

Speaker Name: 

Muthian Sivathanu

Speaker Institution: 

Microsoft Research India

Cookies: 

Yes

Cookies Location: 

1240 CS

Description: 

Can we speed up a CPU-bound server application that does complex in-memory processing under tight latency budget, by moving most of its in-memory state to Flash storage / SSD (which is ~500x slower than RAM) ? Can we achieve crash-tolerance via online process-pair replication and consistent virtual memory snapshotting in a CPU-bound system, while incurring negligible common-case performance hit? Can we speed up petabyte-scale “big data” query processing in a cluster by orders of magnitude while simultaneously reducing cost?

Common knowledge and years of systems research indicate that the above goals are nearly impossible to attain in the general case. So understandably, when faced with a problem of this kind in the context of building a specific system, systems designers typically apply “common belief” and rule out such solutions.

Drawing on experience building large scale production systems, I will give examples where such seemingly “hard” problems were solved by adopting a specialized approach to system design. The present world of aggregated computing - in huge data centers - enables and sometimes justifies building end-to-end specialized, co-designed systems, that achieve significantly better efficiency than traditional approaches. I will present evidence to suggest that systems designers and researchers must use abstraction to empower, not limit, design and be open to challenging/reinventing traditional systems abstractions where sensible, in order to tap into this huge potential in an increasing set of large profitable “applications” such as search, big data, and machine learning/AI.

Bio: Muthian Sivathanu is a principal researcher at Microsoft Research India, specializing in distributed systems, operating systems, and systems for machine learning and information retrieval. Prior to joining Microsoft Research, he worked at Google for about 10 years, mostly focusing on building key infrastructure powering Google web search (in particular, the query engine for web search). He is a UW alum – he obtained his PhD from UW Madison in 2005 advised by Prof. Andrea Arpaci-Dusseau and Prof. Remzi Arpaci-Dusseau. Before that, he obtained a BE from CEG Anna University in 2000.