Advanced and Highly Reliable Distributed Computing in a Fully Automated Environment
Gerd Ballhorn, Martin Bloecker, Danny Keogan, Peter Moberts · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2002
Within the past decades data file sizes and the related computing power for mask data preparation grew linearly following Moore’s law. However, within the last two years the balance between rising data complexity and computing equipment became unstable due to the massive introduction of OPC and the broad rollout of complex variable shaped beam (VSB) data formats. The disturbance of the former linear coherence led to exploding data conversion times (exceeding 100 hours for a single layer) accompanied by heavily escalating data volumes. A very promising way out of that dilemma is the recently announced introduction of distributed job processing within the mask data processing flow. This way was initially introduced to fracture flat jobs. Building on our first promising results last year we now implemented a fully automated design flow with an integrated Linux based cluster for distributed processing. The cluster solution is built in an automated environment in coexistence with our conventional SUN servers. We implemented a highly reliable DP flow on a large scale base which became as stable as our former Solaris SUN system. In the meanwhile we reached a job first time success rate exceeding 99%. After reaching a very stable state we recently started to extend our flat processing conversion steps by investigating hierarchical distributed processing in CATS version 23. We also report on benchmark results comparing new promising hardware configurations to further improve the cluster performance.