Riding The Upgrade Cycle
October 8, 2018 Timothy Prickett Morgan
In an ideal world, all servers would always be new and all operating systems and applications would be patched and running in optimal form. But we don’t live in that world. And that means for many customers that don’t have bales of money sitting around, sometimes the IBM i platform starts to get a little long in the tooth.
That is certainly the case for parts of the IBM i installed base. There are still a lot of Power5 and Power5+ customers out there who have resisted the temptation to upgrade their systems for the past five or six generations, which is a very long time in the computer business – a decade or more.
Many of these companies are technology laggards that are perfectly happy to hang back and get a second-hand machine every time; it’s “new” to them. Still others buy a new machine every four to seven years, swapping out the old box that has very little that can be moved to the new box. And yet others change up their machines more frequently, and for those lucky few using machines with four sockets or more, they have been generally able to do an actual upgrade – meaning keeping the serial number in place so as to not interrupt the depreciation schedule already ticking away on the capital expense and actually moving a new processor and perhaps new memory or other storage into the machine.
The funny thing is that we all talk about upgrades when in fact either model or MES upgrades (they are slightly different) have been done away with in most of the Power Systems line for a while. What companies are really doing most of the time is migrating from one machine to another. IBM does sometime preserve components in the midrange and high-end machines and there is sufficient content that remains the same in the before and after machines that, for accounting purposes, it can be considered an upgrade instead of a new machine and therefore not violate the depreciation schedule. But in most cases, for most customers who hang back for so long, even were those upgrades possible, they might have to jump from one machine to another by hitting one or two machines in the middle, which is a pain in the neck. This is why most customers don’t actually upgrade their systems, in the technical sense, even if they do port their data and applications from one old machine to a newer or the newest one as part of a migration.
Back in the day of the AS/400, this was not the case. While IBM updated the processors every year or so for the entry, midrange, and high-end machines, the memory and other parts of the systems, particularly the I/O, did not change at the same pace and even small machines could be upgraded if customers had a need for more performance. Back then, incremental performance was done in baby steps, but we are now to the point where a jump from a new chip manufacturing process coupled to architectural enhancements results in 2X performance gains between generations, not the 20 percent to 30 percent gains that were typical when IBM was riding up clock speeds and memory capacity in the late 1980s and through the 1990s.
At this time, according to the business partners that I have talked to, most of the deals that customers are doing are to move from Power5, Power5+, Power7, or Power7+ iron, and it is a fairly even split between companies going to Power8 or Power9 systems. You would think that everyone would want a Power9 machine, since it is the latest-greatest thing, but this is not always the case. They may be on an older release that is not supported on the shiny iron, so they have to hang back until their own programmers or those of their third-party software supplier. In many cases, moving to new operating systems triggers an application upgrade event, which triggers an increase in software licensing costs, which means that companies are compelled to hang back even more. The perpetual licensing of software and capital expense of hardware is what makes installed bases get old, for the most part. This is the only way hardware makers and their resellers and software houses can make money, unless they switch to a subscription model for hardware and software that allows them to charge the costs over time. That locks up assets on their balance sheets, rather than on those of the customers, and you have to be a pretty big player with pretty good paying customers to play that game. Which is why the pricing model in the midrange has not changed much in four decades. Leasing and financing this stuff helps with the capital on the balance sheet, but it does nothing about absorbing the pain of changing the physical platform. Even leases that have upgrade provisions built in, which mitigates economic risk to a certain extent, disrupt the operations during the course of an upgrade or migration, which contains operational risks that can lead to other economic risks.
No wonder companies want to stay put as long as they can.
It is even more complicated than it looks, with so many peripherals changing across the generations. You can’t hook two machines with the same serial number to the same Hardware Management Controller, but there are ways to do side-by-side upgrades – under RPQ 8A2356 – where both the old and new machines are allowed to run at the same time with one license for up to 90 days while an upgrade (well, a migration) is performed. For high-end customers, IBM has Enterprise Pools, a kind of group licensing of cores, that allows customers to activate cores on a new machine and use Live Partition Mobility to move things over piecemeal so long as the core pool that was licensed is not surpassed.
There are a lot of things to consider, but given the technical and economic benefits of newer iron compared to older iron, a good case can be made for getting current, particularly for companies that actually have hardware and software maintenance on their systems.
The performance angle is easy enough to argue. If you ignore some pretty important details and look at top running parts, and skip the “plus” generations of chips, you can make some generalizations about the performance bumps that can be expected on a per socket basis for the Power line of chips dating back to 2004. Here is a handy little table you can use as a rule of thumb:
It all comes down to cases, so don’t actually do price/performance and capacity planning based on that table. It is meant to be illustrative, not precise and binding. This is for throughput performance, not for single-threaded job performance, so you have to be doubly careful. (So database workloads versus batch jobs.) Customers with old Power5 and Power6 systems probably won’t jump all the way to Power9, unless they are planning to keep that Power9 machine for a long time and IBM gives them a great deal that beats what a reseller will take for a Power8 machine sitting in the barn. The other thing is that a customer who as on a Power7 machine and who can run IBM i 7.3 might sit this one out and wait to move to Power10 iron in 2020, getting a decade out of the Power7 machine if they think it can make it in the field that long.
I am trying to get my hands on some upgrade (I mean migration) comparisons for Power5, Power6, and Power7 iron, fully burdened with power, cooling, maintenance, and capital costs to help you make your cases for upgrades (I mean migrations) to Power8 or Power9 iron. This really comes down to hard numbers and cash. Stay tuned.