Sun Microsystems sees a future where developers will be able to write tera- or peta-scale applications just as easily as they write applications for just dozens or perhaps hundreds of processors today. Where rich bandwidth, low latencies, very high levels of fault tolerance, and a highly integrated toolset allow researchers to focus on “scaling the program and not the programmers.”
Not long ago, a computational scientist could personally write, debug and optimize code to run on a leadership class high performance computing system without the help of others. Today, things are much harder: the programming for a cluster of machines is significantly more difficult than traditional programming, and the scale of the machines and problems has increased more than 1,000 times. Also, simply owning and running high-end computational facilities for nuclear research, seismic modeling, gene sequencing or business intelligence, takes sizeable investment in terms of staffing, procurement and operations. The complexities associated with HPC continue to increase, and as a result, many advances and scientific discoveries are hampered. For organizations that can afford to staff a sizeable team, it is often the case that the resulting application achieves only 5 to 10 percent of the theoretical peak performance of the system. Often, applications must be restarted from scratch every time a hardware or software failure interrupts the job. The trend toward diminishing productivity associated with coding, debugging, optimizing, modifying, over-provisioning hardware, and even just simply running high-end applications is alarming.
To fill this high-end technology and capability gap, protect critical national security missions, and ensure a new generation of economically viable systems, the United States' Defense Advanced Research Projects Agency (DARPA) has set some very demanding goals for the High Productivity Computing Systems (HPCS) program. By the end of this decade, they've asked for huge leaps such as improving real versus peak application performance by a factor of 10x to 40x, and reducing cost and time for developing solutions by 10x.
Into the Future
Sun's vision of HPC aligns well to meet the needs of the U.S. government and the greater industrial community. Our vision includes systems scaled from thousands to tens of thousands of processors working in an efficient, simple and highly resilient manner. These systems would be able to churn out results that will help lead to new discoveries and provide competitive advantages with relatively little manpower or exceptional programming expertise required, using open source software tools developed by the community.
Impossible? We don't think so. As a Phase II participant in DARPA's HPCS program, Sun has put together an amazing team of engineers and innovators. Led by Sun Fellow and vice-president Jim Mitchell, our “Hero” program (which got its name when Sun Fellow Ivan Sutherland commented that we are undertaking “an effort to build a system of heroic proportions”) has been heads-down designing this revolutionary leap forward in productivity.
Over the last four years we have developed fully integrated system designs based on innovative new hardware and software technologies that we are confident can indeed make this leap. With an emphasis on delivering high levels of productivity to the developer, the system administrator and the facility operator, our research has led us to appreciate the value that massive bandwidth brings to the table – value that translates to increased productivity. Enabling features include globally addressable memory, system level and application checkpointing in combination with hardware and software telemetry for dramatically improved fault tolerance. Advanced features such as these make the system appear more like a flat memory system and allow the developer to focus on solving the problem at hand rather than making elaborate efforts to distribute data in a robust manner.
To achieve our goals, our Very Large Scale Integration (VLSI) Research Group at Sun Labs has been working on innovative technologies to radically improve the bandwidth and latencies associated with chip-to-chip communications. One technology we are looking at is capacitive coupling, which enables high-speed data communication between neighboring chips without the need for wires of any kind. This technology, which we call Proximity Communication, allows for the alignment of metal plates on one chip with metal plates on a neighboring chip and the transfer of data between them with reduced power and with bandwidths and latencies approaching those in native on-silicon communications. The result is comparable to wafer scale integration, but is accomplished by aligning together many small, tested chips. Connecting these chips using Proximity Communication not only reduces system latency but also improves cross-section bandwidth and communication power. A critical area of Sun's work in this area has been the design of system architectures that can best capitalize on this technology.
The ability to connect chips in this fashion is only one part of the bandwidth solution, however. In order to break out of the physical limitations of the X/Y dimension into another plane, we have added breakthrough optical communications technology to the mix.
In order to extend the performance derived from Proximity Communication to neighboring modules and racks, Sun has been working with start-up Luxtera, Inc., a fabless semiconductor company and leader in silicon photonics. Through this partnership with Luxtera, Sun's photonics research team has taken a first step toward a major breakthrough in future optical interconnect for computer architectures. Their design offers massive scalability with ultra-low latency, while still retaining the reliability and cost structure of standard CMOS fabrication.
Within the CMOS circuitry, light enters and is converted from photons into electrons. The data can then be routed to processor, memory controller, I/O unit, graphics engine or other computer subsystems. Data can then be converted back into photons and transported to neighboring nodes, racks, systems, or remote location. This integration of modulators, filters and lasers enables an optical connection to a distant chip via a fiber that can simultaneously carry many wavelengths of light. Future versions could potentially allow data transfer on a single fiber-optic strand to reach 100 gigabits per second, and eventually making it possible to transfer data at a rate of more than 10 terabits per second. In November 2005 at the Supercomputing conference, Sun and Luxtera publicly demonstrated a CMOS nanophotonic link using four-wave dense wavelength-division multiplexing (DWDM) for a bidirectional 40-Gbit/second link. This represents just the beginning of what could result in low-cost, mass-producible, high-bandwidth system interconnects.
Another piece of the puzzle to achieve massive scalability and extreme performance is in the design of emerging object-based storage. Sun's file system research team has made excellent progress working on this technology. By managing data on disk as related data objects rather than unrelated blocks, we are building “smart” storage systems that can self-manage the data they hold.
Object storage file systems delegate space management to the object storage devices (OSDs), which means OSDs have knowledge of the data objects. OSDs can now effectively manage mixed access types by associating accesses with objects, enabling better caching and even pre-fetching. Objects are different than traditional blocks in that they contain both application data and attributes about that data. This enabling technology is the key to supporting a broad range of quality-of-service (QoS) policies such as information lifecycle management and performance guarantees.
The object-based storage T10/1355-D standard (OSD revision 1) was ratified by ANSI in September 2004. Sun is collaborating with Seagate on this standard and is currently demonstrating Shared QFS communicating with a prototype Seagate OSD drive. We plan to include an OSD driver stack in OpenSolaris later this year and are also contributing to the IETF draft standard for object storage pNFS, which enables clients to access data directly from the OSDs, improving scaling over NFS by separating the metadata and data paths. Sun plans to demonstrate interoperability with other vendors' implementations of pNFS later this year.
Fortress Programming Language
Another very exciting area of research is in the development of a new programming language for HPC. We call it Fortress, and our hope is that Fortress will do for Fortran what Java has done for C.
Programming language notation is currently different from the working notations of mathematicians and scientists. Why can't we bring them closer together? Sun is making great progress toward using the dynamic compilation ideas from the Java HotSpot compiler to provide a productivity boost. Essentially, programmers shouldn't have to worry too much about optimizing while they're writing programs. Instead, that optimization can be done by compilers, either ahead of time or on the fly.
Fortress will allow programmers to write more robust code, and will contain more built-in safeguards against error. For example, the root cause of NASA's Mars Climate Orbiter vehicle failure in 1999 was a very simple bug in the algorithm. The programmers simply forgot to translate metric measures into inches, causing the unit to make fatal computation errors. Programmers using Fortress, on the other hand, can depend on built-in intelligence like unit specification: variable X represents feet, Y represents meters, thus X cannot be compared to Y without performing a conversion first.
Sun has published several work-in-progress drafts of the Fortress specification for public comment on its research web site at http://research.sun.com/projects/plrg .
Innovation Leads to New Opportunities
With excellent progress on hypervisor technology, and through leveraging key features in the Solaris 10 Operating System (OS), such as zones and containers, we see the opportunity for massive scaling of the Solaris OS and the system software without massive rewriting.
Other areas of advancement include:
- developing a system architecture that allows a customer to “right size” the amount of performance and fault tolerance they need to maximize productivity for a given budget;
- possible SPARC processor enhancements to meet intensive floating point performance goals;
- automatic parallelization of Fortran giving comparable performance using only 1/10 the source code;
- developer's toolsets that we hope will encourage industry support for an open source HPC toolset;
- an enhanced administrative environment that uses abstractions and automation to enable scalable operations for thousands of resources;
- a visualization architecture to deliver highly scaled performance for huge dataset sizes;
- development of applications using interval arithmetic to ensure accurate results, and in some cases interval solutions to nonlinear problems that are otherwise numerically impossible to solve; and
productivity and performance analysis, simulations, benchmarking and modeling.
The Future — Ultrascale Computing
Sun appreciates the unique opportunity granted by DARPA to do clean sheet design, to explore and research some technologies that just wouldn't have been possible without its HPCS program. It has been a very exciting and challenging few years. We've been very aggressive in the goals we've set – not satisfied to take evolutionary steps into the future, but have set out to develop dramatic game-changing innovation. Sun is happy and proud of the results to date. We believe our Hero technology can change the HPC productivity game; can in fact make it possible for scientists and researchers to solve their most complex and difficult problems.
Mike Vildibill is director in Sun's Scalable Systems Group responsible for next generation server performance and leads product planning for Sun's HPCS program. Mike came to Sun after over a decade at the San Diego Supercomputer Center where he was director of high-end computing. Mike made the HPCwire “Top People to Watch” list in 2002 and continues his passion for performance at Sun.