My PhD research, supervised by Andrew Moore, is concerned with the application of photonic switching technologies to the I/O subsystem of PC scale machines, with a view to improving internal system latency and overcoming the signalling rate (i.e., bandwidth) restrictions which encumber electronic transmission lines manufactured in standard FR4 printed circuit boards.
Photonic technologies may also address the unbridled rate of increase in I/O subsystem power consumption seen over the past decade.
Why does power matter?
Energy is expensive - and it is set to become more expensive. Having consumed it, waste heat must then be disposed of, which requires more energy.
Hot-spots limit local power dissipation in electronic integrated circuits, and temperature is an important factor in the tuning of micro-ring resonators.
Why does bandwidth matter?
Certain scientific applications and financial systems such as the NYSE trading infrastructure have an insatiable need for more bandwidth. Some believe that the rise of manycore demands more bandwidth (although, some don't).
In any case, satisfaction with the status quo is unheard of - yet interconnects such as PCI Express are about to hit the wall like CPUs did a few years ago.
Why does latency matter?
Bandwidth undeniably is important, but what matters more is latency. John Shalf eloquently made the point in his presentation linked above, that latency is what really drives (or limits) real performance.
The HPC folk have known this for a long time, and have invested much effort in low latency interconnects, but...
Why does any of this matter in PC scale machines?
At the moment, I am concentrating my research on photonic interconnects as applied to PC architecture because efforts to build effective photonic switch fabrics at the scale of clusters have been frustrated by the lack of optical RAM needed for buffering. The trade-offs and opportunities involved in PC scale systems are very different. Photonic switching may succeed in PCs where in larger systems they have proved infeasible.
PC scale machines form the vast bulk of the planet's computing resources. They are used in everything from games to high finance. Andrew Bach of the NYSE gave an excellent executive plenary address at Hot Interconnects 16, in which he made the case for - yes, more bandwidth - but mostly less latency, where microseconds mean millions.
Research into the I/O performance of PC scale machines is worth doing because PCs are ubiquitous and very significant to the global economy.
I invite discussion on the subject. Please do email me your thoughts.
Last modified: Fri Aug 29 20:37:23 -0700 2008
$Id: research.html 314 2009-02-08 18:17:38Z djm202 $