Research Interests
I am primarily interested in issues related to memory and storage design. In the recent past, my work has revolved around multiple aspects of system design - from workload characterization, all the way up to proposing and evaluating system-level architectures, with focus on memory and storage sub-system design. Currently, I am interested in the following broad research themes.
- Heterogeneous main memory architectures for smartphones and servers
- Storage system design and optimization, including the system software stack
- Workload characterization of emerging applications, across the entire spectrum of computing devices
In the past, I have looked at memory hierarchy design for Chip Multi Processors (CMPs), at all levels of the hierarchy. More specifically, the techniques to efficiently manage data and capacity in large, last-level caches and multi-socket, single-board NUMA main memories. I have also proposed techniques to manage resistance drift in phase-change memory (PCM) multi-level cells (MLCs), reduce timing and thermal issues in 3D stacked clustered architectures and designed commit algorithms for hardware transactional memory architectures.
Research Projects
1. Non Volatile Memory Architectures
The increase in the amount of data to be processed with increasingly strict latency requirements has made the architecture of main memory a renewed focus area. The main memory, especially on enterprise or datacenter servers is expected to have high bandwidth, large capacity, low latency and low energy consumption. Many of these are contradictory design objectives, and cannot be met only by DRAM, or it’s descendants. However, the advent of new memory technologies has made it possible to design architectures that can provide most of the above requirements. These include emergence of non-volatile memory technologies (NVM) like Phase Change Memory (PCM) and Spin Torque Transfer Memory (STT-RAM), which are being touted as potential replacements for DRAM as main memory. 3D stacking of DRAM, which has also been made economically feasible recently has also led to solutions like High Bandwidth Memory (HBM), and has the potential to be applied to NVMs as well. Despite the availability of architectural possibilities, there is little consensus on which combination of technologies will result in architectures that are best suited for a set of workloads. In this project, we are exploring multiple ways to design better memory hierarchies using a combination of NVMs, traditional architectural optimizations, as well as using novel high performance, energy efficient memory cells for use in caches and main memory.
1.1 Prefetcher Designs for Hybrid Memory Architectures
Hybrid main memory hierarchies are being designed by using a combination of DRAM and an NVM. Typical designs are either splitting the memory address across two or more memory technologies, or using a faster technology with higher lifetimes, typically the DRAM, as a cache for the higher capacity, albeit slower main memory made up of a NVM.
Design of such hybrid architectures remains an active area of research from the perspective of DRAM-as-a-cache design, since DRAM could quickly become the bottleneck, as cache lookups require multiple accesses for reading tag and data. In this work, we are exploring if augmenting the DRAM-as-a-cache model with a novel DRAM cache prefetcher which builds on state of the art Alloy Cache could yield performance benefits. The new DRAM cache architecture allows for prefetching data at both cacheline and page granularities from the NVM, and as a result, provides upto a maximum of 2× performance improvement over a state of the art baseline. (HotStorage ‘20)
1.2 Hybrid On-Chip Memory Hierachies
In this project, we are exploring a “full-stack” solution to designing high capacity and low latency on-chip cache hierarchies by starting at the circuit level of the hardware design stack. We have designed a novel Gain Cell (GC) using FDSOI. The GC has several desirable characteristics, including ~50% higher storage density and ~50% lower dynamic energy as compared to the traditional 6T SRAM, even after accounting for peripheral circuit overheads.
We believe that compared to 6T SRAM, for a given area budget, GC based caches will provide large gains in IPC as well as energy consumption for single- and multi-programmed workloads, respectively, which is being substantiated by our initial results as well.
2. Workload Characterization and Tool Development
To design efficient architectures for the future, we must first understand the behavior of emerging applications on contemporary architectures. Both parts of the equation are equally important to address. If we do not study contemporary (let alone emerging) applications, the chances are that architectures designed using these applications as benchmarks will not be optimized for future workloads in the best case, or will be optimized for workloads that are quickly becoming extinct in the worst. Similarly, studying emerging workloads on near obsolescence architectures fails to reveal system bottlenecks for these workloads.
Workload characterization is a great way to understand the behavior of contemporary applications, and narrow down the sources of inefficiencies and performance bottlenecks in contemporary architectures. This provides us with a set of problems that are worth solving in system design in the near term, as well as provides insights into issues that might become bottlenecks for emerging applications. In this project, carry out workload characterization of emerging applications with a focus on revealing performance bottlenecks in memory and storage systems.
As a part of this project, we have carried out extensive characterization of the SPEC 2017 CPU suite to understand its memory behavior. We have also explored if statistical sampling techniques like SimPoint remain effective for contemporary workloads. In this project, we observed that for cases like memory hierarchy explorations, SimPoints should be used judiciously and with extreme caution in order to derive correct conclusions – inappropriately chosen SimPoint configurations can show large deviations in memory hierarchy behavior as compared to full runs, as reported by prior studies. We have also studied performance bottlenecks in Android, and are currently researching bottlenecks in ML applications using MLPerf. (Bench ‘21, ICPE ‘19, IISWC ‘19)
3. In Memory Compute Architectures
Applications which process extremely large data sets, require large scale movement of data across multiple levels of the memory hierarchy. First, from storage to main memory, and from there to on-chip caches. Finally, the cached data is moved to the CPU registers, where it can be processed. By most estimates, this data movement incurs large latency and consumes a significant amount of energy. Not only that, it causes applications to be bottlenecked by the memory and storage hierarchy. As a result, the need of the hour is for innovations in system design along two major axes. First, the memory and storage hierarchies of existing architectures have to be strengthened to reduce the energy and latency cost of data movement. In parallel, we have to devise architectures that consider data, and not instructions as first class components of computer system design. This would require a diversion from the traditional von-Neumann computing model and move instructions (compute) to where the data is located, than the other way around. We are trying to address both of these important problems in this project via a number of innovative architectural designs for memory hierarchies, as well as exploring non-traditional architectures like in-memory computing.
4. META
The overall goal of the project is to design tools to efficiently explore hybrid memory hierarchies for smartphones and other handheld devices. To that end, we are developing a trace-based simulator based on the Android Open Source Project (AOSP) which can simulate multiple types of memory hierarchies, including ones that are architected using emerging memory technologies, as well as arbitrarily deep cache hierarchies. The tool will also be capable of executing application traces across multiple Android versions. (DATE 2018, MobiCom 2018)