Last year several of the Georgia Tech Systems professors from CS and ECE banded together to form the Center for Experimental Research in Computer Systems (CERCS). We've been lucky to have sponsors such as Intel fund our work and provide us with multiple clusters for our experiments. I helped out with the below poster about Active SANs work.
After getting GRIM up and running, I did a lot of work hooking it up to differerent peripheral devices. For the Novel uses of SANs workshop I wrote a paper that descrives how to build a GRIM communication endpoint for FPGA accelerators, and how users could set up a computational pipeline for the devices with some simple forwarding tables.
This paper explores the view that the SAN network infrastructure can be an active computational entity capable of supporting certain classes of data intensive computations effectively during communication. The performance is achieved via the use of Field Programmable Gate Arrays (FPGAs) in the network interfaces (NIs). This paper describes the programming model and the design of a prototype hardware/software implementation using commercial FPGA devices coupled with Myrinet. An active messages style of programming is used to support application-transparent, dynamic reconfiguration of the FPGA hardware to accommodate different computations over time. Performance evaluation of this implementation quantifies the overheads and sources of performance improvement.
This paper I wrote for the first Myrinet User Group meeting provides an example of how GRIM interacts with an I2O peripheral device. It also goes into the details of how the NI uses some cut-through optimizations to boost the bandwidth for sends in the system.
Resource rich clusters are an emerging category of clusters of workstations where cluster nodes comprise of modern CPUs as well as high-performance peripheral devices such as intelligent I/O interfaces, active disks, and capture devices that directly access the network. These clusters target specific applications such as digital libraries, web servers, and multimedia kiosks. We argue that such clusters benefit from a re-examination of the design of the message layer to retain high performance communication while facilitating the interface to endpoints for a variety of devices.
This paper describes a message layer design which includes optimistic flow control, the use of logical channels, a push-style cut-through injection optimization, and an API supporting cluster-wide active message handler management. The goal is to support a number of diverse cluster hardware configurations where communication endpoints exist in a variety of locations within a node. The current implementation has been tested on a Myrinet cluster with communication endpoints located in the host CPUs as well as Intel i960 based I2O server cards.
The initial three years of my PhD were funded through a NASA Graduate Student Research Program fellowship. One of the benefits of this program was that they encouraged students to come out during the summers to learn more about the problems that NASA faces. I spent my 1999 and 2000 summers in Pasadena, working in the Center for Integrated Space Microsystems. While I initially had plans to evaluate how well my GRIM software worked with one of their clusters, my center director asked if I'd be interested in trying something a little different. He told me that the recent success with the Sojourner rover had sparked a great deal of interest in deploying more in situ sensors on Mars. He challenged me to think about engineering problems NASA would face if it were to cast hundreds to thousands of wireless sensor nodes across Mars.
It was very different than the PhD topic I had been studying, but it was too interesting to pass up. I dove into the papers to learn what people had been doing with wireless sensor networks on Earth. As a networking person, the part that interested me the most was figuring out how a collection of low power, low performance sensor nodes would boot up, establish a routable network, and then collect meaningful information over a geographic region. There were many examples to draw from on Earth, including battlefield sensors, buoy networks, and arctic tumbleweed sensors. My director introduced me to people from all over the lab to learn more about how NASA builds resilient embedded systems that are designed to survive being dropped out of the atmosphere into an environment with harsh thermal constraints.
While the NASA summers took me off course into a side topic that delayed my graduation, it was one of the best things I did during my academic career because it encouraged me to think about hard problems that were outside of my comfort zone. I wrote up the below technical report summarizing some of the things I learned, though I never got it officially entered for a report number at JPL or Georgia Tech. I put the paper up on my school web page, which would up getting referenced more than I would have thought.
This is the first paper where I talk about the General-purpose Reliable In-order Messages (GRIM) communication layer that I've been developing for Myrinet. My idea is that we should be building communication libraries that can route data between host processors and remote peripheral devices that are distributed throughout the cluster. We're calling these types of systems "resource rich clusters" because they have more computing resources than other types of systems. The current version of GRIM offloads a lot of the message layer management responsibilities to the network interface, which simplifies the amount of effort required for peripheral devices to communicate in the platform.
Resource rich clusters are an emerging category of computational platform where cluster nodes have both CPUs as well as high-performance I/O cards. These clusters target specific applications such as digital libraries, web servers, and multimedia kiosks. The presence of communication endpoints at locations other than the host CPU requires a re-examination of how middleware for these clusters should be constructed.
A key issue of middleware design is the management of flow control for the reliable delivery of messages. We propose using a network interface based optimistic flow control scheme to address resource rich cluster requirements. We implement this functionality with a message layer called GRIM, and compare its general performance to other well-known message layers. This implementation suggests that the necessary middleware functionality can not only be constructed efficiently, but also in a way that provides additional middleware benefits.