EDA companies should rethink their attitudes to cloud computing to spur a new phase of industry growth, according to an Intel software architecture manager who has been involved in a project to improve the utilisation of cloud resources for EDA jobs by up to 15%.
Speaking at a DAC poster session on Tuesday, Naresh Sehgal said: “We need the EDA companies to take some risk and and rent their tools by the hour.”
He said a colleague at a start-up had wanted to use commercial placement tools to help optimise the layout of a new hospital for health insurers Keyser Permanente, but couldn’t afford the upfront licensing fee for the tools.
“If you made the software available cheaply for rent people would try it out in different domains. The EDA industry has to think differently.”
The poster, co-authored by James Colgan of cloud company Xuropa, Sehgal and Mrittika Ganguli of Intel Bangalore, discusses research on improving the utilisation of cloud resources for EDA jobs by developing a better understanding of the ways they use the underlying processors.
The problem with cloud-computing strategies is that the virtualisation mechanisms which abstract processing jobs from the underlying processing power can also obscure the details of the way that the job is consuming those resources. The tool that the project team have developed monitors aspects of processor utilisation such as overall load, I/O, cache use, network bandwidth and memory access and relates that information to the virtual machine that is running on the processor. This information is then used to improve the scheduling of new simulation jobs, so that, for example, memory-intensive tasks are not scheduled on processors running other jobs that are already consuming a lot of memory bandwidth.
“When people do scheduling today they are leaving a lot of money on the table,” said Sehgal. The monitoring scheme has now been submitted as a proposal to OpenStack.org, a project to build an open-source software architecture for public and private clouds.
James Colgan, founder and CEO of Xuropa, which built the cloud environment that enabled Cadence to run its tools in the cloud without having to worry about access control, security and licensing issues, said that the fundamental reasons why semiconductor vendors aren’t rushing to use the cloud for their work are two-fold: “EDA vendors haven’t worked out how to deliver tools to the cloud in a way that is accretive. The second issue is what happens if something goes wrong and it becomes an issue of liability for data loss or corruption. That’s where life gets very sticky.”
Sehgal pointed out that people used to keep their money under their mattresses because they didn’t trust banks, but eventually bank processes improved and trust was established. The same could be true for cloud-based EDA. He added that it could also enable new approaches to chip design.
“There’s no reason you couldn’t go to the cloud using tools you are renting by the hour and then pass the design along to a fabless semiconductor company, so that the first time you saw anything physical was when they put a chip in your hand.”
Colgan argued that it should be possible to go a step beyond the fabless company model to CADless companies, where the tool licences are held in the cloud not on the premises.
“Could this happen? It all comes down to the economic imperative,” he said. “If you’re like Intel sitting on 70,000 servers there is no reason to go the cloud. For small companies, though, this could be a differentiator.”
How would they get the expertise to run large jobs using rented licences on a cloud infrastructure? Part of Xuropa’s role in they research project was to work with the Intel team to configure and debug the cloud environment, and then to write scripts to run the Cadence simulation and verification tools as efficiently as possible, resulting in the 15% performance gain.