by Tets Maniwa

Continued from Part 1

April 1, 2010—SNUG Santa Clara, CA—A continuing review of a panel at the Synopys User Group (SNUG) discussing cloud computing and EDA. The panel participants were Hasmukh Ranjan of Synopsys who moderated the panel, Scott Clark, director of engineering for information technologies at Broadcom, Kishore Singal scientist and IT manager for Synopsys, Jeff Barr, senior Web services evangelist at Amazon, and Vijay Bollapragda, manager of engineering infrastructure at Cisco Systems.

The semiconductor industry is in a period of transition. A major concern hindering adoption of cloud computing is security. At the same time, the industry is looking for ways to manage design costs and complexity as well as tool infrastructure cost and complexity.

Singal noted that high performance computing is moving in two directions, high performance, multi-core processors and larger arrays of highly distributed and low-cost computers. The challenges for EDA users include data security (already partially solved and implemented in other industries), loss of control, Internet bandwidth, and performance degradation due to server virtualization.

Despite these challenges, some applications which are amenable to a cloud computing include software regression, training, and demonstration and evaluation. These are all applications with relatively small I/O and relatively light computing requirements. Some design tasks which might move to clouds and are Monte Carlo simulations, corner simulations, sweeps, and some verification. More challenging applications have large I/O requirements as well as high levels of computation. Some of these larger applications are place and route, full chip extraction, highly interactive graphics such as layout editing and manual schematic entry.

High-performance computing and general-purpose GPU are highly threaded applications that are directed mostly towards games and need to be as fast as possible at the lowest possible cost. A likely scenario would be to have a hybrid approach with a mix of both private and commercial compute clouds linked via a tunnel. By developing a common infrastructure at both ends of the pipe, users and compute vendors could develop controllers to distribute workloads across the public and internal server farms.

Most EDA and cloud computer vendors must develop the capabilities to identify resource requirements for application so appropriate memory, computer, and storage capabilities are available for the jobs. In addition, EDA tools must adapt to this new environment and migrate from very large data sets running on high performance computers to smaller data sets running on a cluster of lesser configured machines.

Barr described Amazon's web services as a collection of computers and storage series of complexes with the ability to supply on-demand, pay-as-you-go capabilities that a user can access using Amazon's existing infrastructure. They're running a large number of programs from a large variety of users including some functions that are deemed mission critical and/or high-security. They can provide on-demand resources with as little as 30 seconds of activation time.

They can scale resources up or down and can supply computers with Windows, Linux, or open Solaris operating systems. The computers are located in three regions, two in the United States and the third in Europe. They will be opening a new center in Singapore soon. The computer systems are divided into eight independent zones and can be configured with from one to 68 GB of RAM in either 32 or 64 bit configurations. Costs range from $0.085 to $2.45 per hour. Then a number of different pricing plans—on-demand, reserved, and preset which allows you to run computers only when pricing its activation point.

Bollapragda is supporting 25,000 engineers and has to address many back-end issues. To enable cloud computing in his environment, he needs appropriate products, solutions, services for secure computing and storage, and standards so that the underlying technologies become immaterial. The standards will include some type of unified fabric to minimize I/O compatibility issues; unified computer, network, and storage management to facilitate policies and controls; integrated virtual systems to coordinate the services in the data centers; enterprise-class compute resources in the cloud; and finally comparable resources in the service provider cloud.

The challenges Cisco face include static islands of resources and tools that are optimized for peak loads. They have at least 10,000 servers allocated as 2 1/2 to three engineers per server plus some more servers in the labs and on people's desktops. The applications running on servers include EDA, software development, Web servers, databases, test, and all the other miscellaneous functions needed within the company. They have found it takes between two and five years to build a data center, and after it's fully loaded the asset utilization average is 20 percent.

They plan to develop an internal cloud environment and over time migrate some of the functions to some service provider. To this end, they have started to develop standardized service workload profiles to abstract the hardware attributes from any particular machine. They're trying to develop automation tools to manage server allocation and complexity in developed two or three configuration profiles for the majority of their work loads. They are trying to develop just-in-time provisioning at various scale points and have developed methods to abstract the MAC address to be able to move jobs associated with fixed node keys. These techniques will enable them to work loads as needed.

After the presentations, the audience started asking questions. The first ones were about the possibility of getting licenses that were able to use the cloud capabilities. Ranjan gave the noncommittal "we are working on it." Barr suggested running flexLM on one of the cloud machines or running a license server with access to the cloud. Clark asked that EDA vendors base license servers to work with any resources available to that server, including external clouds. He also suggested the best use for cloud capabilities were those that allowed easy data duplication and gave the best representations of the existing data.

A follow up question was how to get EDA vendors to step up to this challenge. The hardware exists but current EDA licenses do not allow for interruption of a job without re-invocation of the license. In the cloud environment, if the run priority changes or CPU changes, how does the user restart the job? Bollapragda joined in and added if they are licensed for peak loads, can the virtualized machines be configured for license on demand. He is still waiting for EDA vendor approvals for this mode of operation.

A question to Amazon's Barr was related to configurations: "why so little memory on the servers", and "what would a fully loaded machine cost?" Barr replied most of their current users used 2-50 GB. He noted that 500GB is a lot of RAM and that costs would depend on configuration, so a machine with lots of RAM and little CPU would cost less than one with lots of memory and lots of CPUs.

Questions of security and levels of encryption brought forth a number of responses. Singhal remarked that IP is already encrypted, but users will also need to compress the data if the normal design files were transferred in and out of the cloud. He suggested a lossy compression like that used in games and TV might work. Clark observed that companies already address data transfers in the model that is used to transfer design data to the foundries. Bollapragda added that data transfer is also an issue, because there are no standards for transferring the very large data files used in design. Current efforts are provider dependent and not interoperable. Standards are only needed for automatic transfers. Manual transfers do not need standards since the user will identify all necessary parameters for the data transfer. Barr injected that no one has identified what type of standards are required. For some customers, large files are shipped on a hard drive to one of the server farms and up-/down-loaded and returned. For very large files, FedEx may be faster than the Internet.

Of course, everyone wanted to know if the suite of tools in current use could ever be available in the cloud. Barr responded that they provide logical images of virtual machines that can be pre-configured as needed. They could preset and pre-configure machines so the user only needed to pay for long-term storage at $0.10 to $0.15 per GB per month. By setting up a working environment beforehand, a user could log in and easily find "his" machine.

Everyone was interested but wanted to know when this capability would be available.
Singal acknowledged that no one is actually doing design work on cloud computers. So far they have not been able to identify many commonalities in work environments. Clark suggested workload management frameworks (which already exist) might be the management key to success. Bollapragda added that the challenges of data and tool dependencies create problems for data exchanges. One path for progress might be to change the tool structures and data models so only sparse data were necessary with detailed data filed in at execution.