newsroom

In my previous blog post, Improving chip yield rates with cognitive manufacturing, I highlighted the costs associated with semiconductor manufacturing, and how cognitive methods can yield benefits in both design and manufacture.

The compute farms and private clouds required to run the necessary simulations and verifications can be pretty substantial, with the largest companies deploying hundreds of thousands of compute cores.  Frequently, the applications used in chip design are primarily sequential and only utilize one core at a time.  And in many cases, the runtime is quite short, which means many potential millions of simulations/verifications per day. Efficiently and intelligently scheduling such volumes of work are well beyond manual capabilities.

Unlike the way it works within some other industries, most applications are licensed from independent software vendors (ISVs) with prices ranging from a few thousand dollars to hundreds of thousands of dollars per instance per year.   Large design companies can end up paying tens millions of dollars per year.

In many cases it is a lack of application license capacity rather than compute capacity that is the limiting factor on time-to-results. With expenditures like that, you need to ensure that you are not only getting high license utilization but that you are also ensuring that the right license is being used for the right person on the right project.

For example, if you are paying $20 million dollars a year in license fees and only getting 50 percent-effective utilization — that is a lot of wasted money.  If you could drive it closer to 100 percent, you could potentially get the same work done, in the same time, with only half the licenses…. well that’s a savings of $10 million dollars right there.

But more importantly, keeping the same number of licenses, you could cut the time needed in half.  Or you could take the same time, but run more simulations and verifications — resulting in a better product.  Or finally, you could pick the optimum combination these characteristics for your business.

IBM Spectrum LSF License Scheduler addresses these challenges.  It enables license sharing between local or global project teams.  Using flexible policies, it ensures that license availability is prioritized by workload, user and project, and that licenses are optimally used instead of simply throttling jobs based on license. Without any form of managed scheduling and allocation, the system is a free-for-all — whoever requests the license first gets it and keeps it.

Figure 1 shows a typical trace of license usage.  The workload scheduler allocates licenses to jobs, and at some point the job actually makes use of it… hopefully.  There may be cases where it doesn’t use that license at all, or uses an entirely different license.  Other licenses are being reserved so urgent work can start as soon as possible – the net result is that utilization suffers, lots of licenses are reserved or allocated, but unused.   That’s a lot of wasted money.

Figure 1

In figure 1 we can see that there is always a certain number of licenses reserved (the red area). A simple improvement would be to enable overallocation – tell the scheduler that there are more licenses available than there really are (purple) – and in figure 2 we see that this can make a noticeable difference.

Figure 2

However, we’re still losing 15-20 percent efficiency as the “reserved” amount is not constant.

By applying cognitive techniques, the system can learn how licenses are really being reserved, allocated and consumed, dynamically adjusting the (purple) buffer in response to changes in reservations (red), thus driving actual license utilization (green) very close to 100 percent.

Figure 3

By taking advantage of these techniques in conjunction with detailed license usage reporting, clients have been able to understand the real usage of application licenses. They have also been able to not only remix their existing license portfolio within their existing budget but also realize significant cost and time to market benefits.

Furthermore, it becomes much easier to justify the investment in additional licenses when you can prove to the CFO that you are using everything you already have effectively.  We’ve now made it easier than ever to try IBM Spectrum LSF and the capabilities described in this paper.  Learn more about IBM Spectrum LSF evaluations here.

The post Cognitive scheduling yields savings and faster time to market appeared first on IBM Systems Blog: In the Making.