Cadence presents plan for piecemeal cloud compute
Cadence Design Systems has launched a web-based EDA service the company hopes will ease the transition from self-hosted computing to more flexible cloud-based development.
Although the CloudBurst service will make all of the company’s tools available online through a web browser, the initial focus is on tools for design analysis and sign-off that need to be distributed across many nodes to complete in a workable timeframe.
Many EDA customers have the problem of trying to maximize the utilization of their self-hosted environments to avoid capital expenditure from becoming too high a financial burden. “The CFO is looking over your shoulder to ensure that capital spend is productive. In order to optimize for utilization, you want a little less capacity than what you truly need,” said Craig Johnson, vice president of cloud business development at Cadence.
The difficulty of absorbing peak-compute demands, particularly during the hectic pre-tapeout period makes cloud-based hosting attractive to many. “With the cloud, the mindset can change to one of maximizing throughput.”
Migration headaches
Using cloud servers for EDA comes with practical problems. Since launching its initial offering [a year ago], Johnson said customers continue to face challenges in migrating their design environments away from self-hosted systems. “What makes it challenging in our space is that EDA is extremely complex. Tools have to work together but no two customers of ours have a flow that is the same. Many have significant investments in their compute environment and the flow is often tightly coupled to their IT solution. The scripts and profiles are tied to that environment. That presents a challenge for adopting the cloud.
“On our own journey to the cloud, we found some of these scripts predated some of the people using them. As we attempted to move jobs to the cloud we would get ‘file not found’ situations,” Johnson explained. But there is strong reluctance to change the scripts because of the fear of introducing subtle bugs that could kill the design.
“You need a way to keep using your private farm at the highest level you can and leverage the public cloud when your compute capacity is insufficient,” Johnson said.
With CloudBurst, although it can support the entire Cadence portfolio, the emphasis is on tools that can be deployed tactically and, in effect, broken out of the standard flow: design files can be copied to the cloud for processing and the results copied back for integration into the rest of the flow. “The approach doesn’t require any IT involvement on the customer side: it’s 100 percent web-based,” Johnson said.
Transfer acceleration
Although the tactical usage of cloud tools helps avoid the problem of breaking scripts in tightly coupled flows, the approach still presented problems in terms of implementation. The biggest was the burden of copying files that can consume more than a terabyte of data. “If you try to move all that data around and you don’t have the right transfer technology, the compute capacity is no use to you,” Johnson said.
The upload and download mechanism is akin to that of BitTorrent. Rather than trying to force the transfer through one connection, the files are broken down into chunks, encrypted and then relayed using multiple machines on the client’s networks. Cadence believes this will make it possible to maximize the throughput of each transfer and make it feasible to move large files in and out of the CloudBurst-based tools. A second technique is to use incremental uploads, although in many cases small changes can lead to large-scale changes in overall file because of dependencies, which will limits the efficiency of any incremental system.
A tool launched alongside CloudBurst provides an example of the kind of workload Cadence expects to host. Clarity is a 3D field solver designed to handle the problem of simulating the electrical interactions that affect high-speed connections through a large system, such as a rack-mounted server or a vehicle chassis. The product is the result of rearchitecting the way finite-element meshes are generated and simulated so that the workload can be divided across hundreds of compute nodes. Brad Griffin, Cadence product-marketing group director, said the distribution can lead to ten-fold speedups for large simulations.
Cadence will offer two options for charging for cloud-compute usage. One is to have all payments going through the EDA supplier: invoices will reflect a combination of tool license usage and a pass-through of the charges from cloud providers themselves. Alternatively, customers who already have supplier agreements with AWS or Azure can pay separately for the compute and only be charged for the license rental by Cadence.
