(Google Scholar: y. charlie hu)
The following 15 selected publications represent a good sampling of our work:
This paper points out that several fundamental assumptions behind power models for mobile devices developed in the 1990's do not hold on modern smartphones, and develops a system-call-based power model -- the first power model that captures unique smartphone power behavior yet can map the phone component power beheavior back to the program entities that triggerd them. The model laid the foundation for developing energy profiling and diagnostic tools for mobile apps on modern smartphones.
Building on a key property of the above power model, that using system
calls as triggers in the power model enables easy mapping of power
activities back to the responsible program entities in the app source
code, the paper develops the first source-code-level energy profiler
for smartphone apps, 30 years after the original performance profiler
gprof from UC Berkeley (published in PLDI 1982).
Eprof enables energy-aware app development --it identifies energy bottlenecks among potentially millions of lines of complex app source code. It has been commercialized by Mobile Enerlytics, LLC, and deployed at top app vendors such as Pandora and Handset vendors such as Lava International to optimize the battery drain of apps running on a few hundred millions smartphones.
This paper conducted ground-breaking research on energy bugs in mobile apps, by characterizing the prevalence of a class of energy bugs in smartphone apps that are caused by misuse of the Android wakelock power control API by app developers and their severe impact on the smartphone battery life, and developing the first detection technique based on static analysis for detecting them.
Two months after the paper was published, Google implemented a version of the algorithm in Android Development Toolkit which has been used by Android app developers all over the world.
Eight years into the smartphone era, with 2 billion smartphones around the world, and about 700 papers following up our original work (above papers), the industry and research community still had no clear understanding of how smartphone battery was drained in the wild under normal usage. This work (1) solved the key technical challenge of how to measure detailed energy drain (by components, by apps, and by activities) on unmodified phones in-the-wild; (2) performed the first large-scale in-the-wild measurement study (1520 Galaxy phones from 56 countries) of energy drain of smartphones under normal daily usage, (3) provided the first insight into major sources of energy drain in-the-wild, which have significant implications to key players of the Android phone eco-system, including phone vendors (e.g., Samsung), Android developers, app developers, and the billions smartphone users, towards the common goal of extending smartphone battery life and improving the user mobile experience.
The study was widely covered in the media such as this (Scientific American podcast.)
The paper developed HUSH, the first app background activity manager inside Android that applies learning to dynamically suppress app background activities learned to be not useful to individual phone users, and was shown to reduce the average daily battery drain by 16% across a 2000-galaxy phone/user trace.
In 2015, Adroid Mashmallow added a feature very similar to HUSH called app-standby. HUSH was released in Github and was forked over 50 times by the Android developer community to retrofit into older versions of Android, and widely covered in media such as BBC.
A long-standing, fundamental research challenge in Distributed Systems had been how to design efficient Distributed Shared Memory systems to support parallel computing on networks of workstations. This work developed the first system that enables efficient and transparent sharing of data with both fine-grained and coarse-grained access patterns, and practically closed the chapter on the 15-year-long intensively-researched topic in Distributed Systems.
These two papers (250 cites) contributed to an OS design trend that exploits architectural information readily available during program execution such as the program counter. The new buffer cache replacement technique developed was cited in Tanenbaum's textbook on Modern Operating Systems (3rd Ed.), and was once pursued by Linux core developers for incorporation into Linux.
All File System buffer cache replacement algorithms proposed since 90’s were studied without taking into account kernel-driven prefetching, supported in all modern operating systems. This "whistle-blowing" paper points out the significant fallacy in this methodology. It shows kernel prefetching can have a significant impact on the relative performance of replacement algorithms. The study had an immediate impact; all subsequent papers on buffer cache replacement techniques incorporated kernel prefetching.
Networking reservation is performed to ensure predictability in application performance in cloud datacenters, but the previous abstraction for specifying application network requirement assumes the requirement remains constant over app execution. The paper for the first time points out application networking requirement is time-varying and proposed the first simple yet effective abstraction that captures such behavior, and shows such an abstraction results in significantly higher utilization of data centers and lower cost to cloud customers than prior art that ignores the time-varying effect.
This work developed the first Traffic Engineering system for Online Service Provider networks to jointly optimize their ISP cost and end-to-end user traffic performance, and was shown to reduce Microsoft OSP's ISP cost by 40% without any increase in service latency.
IP prefix hijacking poses a serious threat to Internet routing security and was intensively studied by the SIGCOMM community and industry. All previous proposals required a large number of monitors distributed over the Internet. The paper shows that a simple, elegant IP prefix hijacking detection system, iSPY, using exactly one monitor in the home ISP, can overcome all the drawbacks of prior art. The work practically closed the chapter on IP prefix hijacking detection.
Scalable routing in mobile ad hoc networks remained arguably the hardest problem faced by mobile ad hoc networking research throughout the 90’s into early 2000's. This work shows that a simple, practical geographic hashing based location service for mobile ad hoc networks can be far more scalable and robust than previous state-of-the-art, highly complex hierarchical location schemes such as GLS from MIT (MobiCom 2000), and influenced all future research on this hot topic.
This paper for the first time points out p2p overlay networks and mobile ad hoc networks share significant commonalities, and advocates exploiting the synergy between the two as a promising new area. The paper stimulated much community research interests and activities, e.g. Dagstuhl Seminar on "P2P MANETs: New Research Issues" in 2005, creation of the annual IEEE Workshop on Mobile Peer-to-Peer. The HotOS and its follow-up full paper (Ekta at WMCSA 2004) were cited over 380 times.
This and a companion paper (SC 2004) started one of the first two significant efforts (the other was by Ian Foster) on exploring the peer-to-peer (p2p) paradigm in solving a key challenge in the Computational Grid: how to manage shared resources over multiple administrative domains, such as compute cycles (this paper) and storage (SC 2004).