Reader Feedback on Sundry Recent Stories
October 19, 2009 Timothy Prickett Morgan
We got some more feedback on the core, thread, and parallelism issue that I brought up in The Four Hundred, which I thought I would share with you. I don’t claim to have all the angles covered, and readers seem to understand that I was just bringing up a potential problem I see on the horizon as the Power7 chips are packed with lots of threads and cores. We also got some feedback on the iManifest effort in the United States and Big Blue’s big bet on the mainframe in April 1964.
Feedback on Moore’s Law and the Performance Wall:
My compliments on another interesting article, Timothy. You probably gave too much credit to Java for its ability to spawn threads. A thread is slightly more lightweight than a process, but in the vast majority of cases, all work performed by threads much be synchronized, which is a big hurdle.
Consider a process like WebSphere Application Server, which spawns threads to evoke servlets. You can configure it to spawn hundreds or even thousands of threads and the job will never be able to fully use more than one or two processor sockets, due to synchronization.
So what do we do on a multi-CPU server? We configure and launch multiple instances of the application server, deploy our applications across each instance, and come up with some external means to balance workloads across multiple instances. That gets really tedious.
Your article points out a reference to RPG’s thread(*serialize) specification, which is RPG’s way of implementing synchronization, when ILE modules are called from multi-threaded servers. But you failed to point out that Java methods must be synchronized as well. They enforce the same types of constraints.
For business applications, the question is which performs better–spawning (and synchronizing) threads within Java application servers, or launching separate ILE *PGMs and bypassing the synchronization hurdle–letting IBM i dispatch application threads to physical processors? My testing and observations favor the latter, by a big margin.
One answer to the problem you raised is in Intel’s recent TurboBoost technology. In heavily threaded environments, no gain. In lower loads/threads, the power envelope allows for a faster single thread/core. That would help RPG apps a lot. There is nothing in TurboBoost that IBM couldn’t copy. After all, it’s all about what you do with the transistors Moore’s Law gives you.
Hope you are well.
True enough. With TurboBoost on a “Nehalem” Xeon 3400, 3500, or 5500 processor, if you shut down three of the four cores in the chips, you can crank the clock speed on the remaining core that is still live. Intel’s implementation of TurboBoost is pretty unimpressive as far as I am concerned, and that is because heat rises on a log scale as clock speeds go up. On regular Xeon 5500 parts, a quad-core chip with a basic 2.93 GHz clock speed can crank one core up to 3.33 GHz and still stay within the 95 watt thermal envelope. Personally, I am not impressed by that 13.7 percent increase in clock speed when you shut down 75 percent of the cores. This is awful, really.
It seems to me that what you really want to be able to do is have a core crank up to 4 GHz or 5 GHz on a Power7 chip if four or six of the eight cores are shut down, if the base speed is somewhere around 3 GHz as I expect. IBM’s implementation of simultaneous multithreading has always been better than Intel’s, and I see no reason why its TurboBoost variant, should one appear in the Power7 chips, be better, too.
For another interesting reader feedback article on the parallelism issue, see Reader Feedback on Moore’s Law and the Performance Wall from last week’s issue.
Feedback on Ellison Wants Oracle to Be IBM 1.5:
Thanks for another great article and for your incredible perspective.
Unfortunately, there is no one at IBM who is listening.
I was an IBMer on April 7, 1964, and I was at the announcement meeting of the System/360.
We all could feel the importance of the announcement and we enjoyed the following glory years of IBM because of it and Tom Watson Jr.’s big gamble.
I wasn’t even an itch in my daddy’s pants then. HA! Wait. My sister was born in May 1964, and I was born in March 1965. OK. So maybe I was an itch. . . .
I hope you are well as well. I am hanging in there.
Good article. I see some bloggers suggesting that iManifest is already out of steam but I hope not.
I like your idea for a grassroots campaign. As I looked at Olen’s Website, I thought, how do I give $25, $50, or even $100 to a cause that is so near and dear to my heart and supporting a community that provided me a 30-year career. I’d rather allocate a few bucks to this than see IBM i just fade away.
Keep pressing this matter and se if they come up with something. Maybe we need a “Walk for i” each year to raise money though it seems unbelievable that it has come to this. Why don’t they (IBM) get it?