Antoine Saint Exupery
I am interested in computer architecture and systems. Specifically I work in improving resource-efficiency in large-scale datacenters by designing practical and scalable systems for cluster management, scheduling and resource management that are QoS-aware and resource-efficient. My approach is that given the fact that traditional techniques to improve the compute capabilities and efficiency of datacenters (e.g., switching to commodity computing, relying on Dennard scaling) have reached the point of diminishing returns, we must now focus on using the existing servers more efficiently. In the past I have also worked on datacenter application modeling. I believe that efficient system designs should have a strong analytical foundation. Here is a list of projects I am currently working on or have worked on in the past.
- Quasar: Resource-Efficient and QoS-Aware Cluster Management Traditionally, datacenters have struggled with increasing utilization, primarily due
to users overprovisioning resource reservations to side-step performance unpredictability.
Quasar acknowledges that determining the correct resources for a job is a difficult, multidimensional problem for users and moves this responsibility to the cluster
manager. In turn the user only needs to specify a performance target for his job. Quasar leverages efficient techniques that find similarities between previous and
new applications to translate this performance target to resources, much like a movie recommendation system finds similarities between previous and new users to
recommend movies that they are likely to enjoy. Quasar achieves both high cluster utilization and high per-application performance.
[paper] [demo] [press]
- Paragon: QoS-Aware Scheduling in Heterogeneous Datacenters Paragon is a QoS-aware scheduler that accounts for both inteference between co-scheduled workloads
and platform heterogeneity when assigning applications to servers. The scheduler leverages fast classification techniques to determine the interference
and heterogeneity preferences of incoming applications, only adding minimal scheduling overheads to the system. Paragon improves system utilization significantly,
while maintaining each aplication's performance requirements and is scalable and lightweight.
- Performance and Bandwidth-Aware Storage Consolidation in Datacenters: Large-scale workloads typically have well-characterized activity patterns. Specifically in storage, we can study these patterns to identify periods of low utilization and consolidate the storage
or multiple applications on fewer storage nodes. We proposed BLOC, a bandwidth-aware scheme that adaptively changes the number of active storage nodes to adjust to user load.
BLOC improves utilization by 47% without degrading performance and is transparent to the applications.
- Datacenter Application Modeling: Previously, I worked on characterizing and modeling the behavior of large-scale datacenter applications. We develop concise analytical models, both for the storage and network part of large-scale workloads
and validate their accuracy against real datacenter applications. These models decouple performing meaningful system studies from the requirements to have access to real large-scale applications.
They enable studies, such as BLOC, that were previously infeasible due to the unavailability of representative datacenter applications. For more information, take a look at our papers on
storage and network workload modeling.
I also enjoy teaching and mentoring students.
- In Fall 2014 I am co-teaching CS316 (Advanced Processor Design).
- In Spring 2014 I was co-teaching EE282 (Computer Architecture).
- In Fall 2013, I mentored several quarter-long projects for CS316 (Advanced Processor Architecture) related to heterogeneous CMP scheduling and datacenter server provisioning.
- In Spring 2013 I was TAing EE282 (Computer Architecture) and teaching some of the lectures and a weekly recitation. Over the years I've mentored several students either as part of quarterly or summer-long projects.