AIX 7.1 moves forward to Power7 iron

You're timelines are wrong.

1. What were you doing in 1990. I would doubt that it was working in the UNIX field.

2. Sun did not have a logical volume manager at that time, unless you count Veritas, which was not their product Sun Disk Suite was a charged for add on that was released in 1991. Nor did HP until sometime after AIX 3.1 was launched (the HP/UX logical volume manager was derived from the code IBM contributed to OSF if I remember correctly, and added to HP/UX 9.0 in about 1992)

3. HP/UX had SAM. Hmm, not good, and not a patch on Smit.

4. I'm sorry, Solaris definitly did not have dynamically loadable AND unloadable device drivers (I'm not talking linking here, I'm talking device drivers) at this time. I was having to sysgen systems to tell it what devices were included in the Kernel. And I would ask when you think that Solaris was launched, bearing in mind that Solaris 1 was a packaging option including Sun OS 4, and was not the label for the entire OS until Solaris 2 which was SVR4 based in around 1992.

5. Sun OS was a BSD derived OS until Solaris 2 so of course it implemented BSD commands. It had a veneer of SystemV on some commands until the SVR4 initiative that was not Sun's idea (remember this). The way that you effectively chose which type of commands and environment you used was poor. You may not think this was important, but as you pointed out, this could just be an opinion.

6. I'll give you Starfire. I had forgotten about that system. As I pointed out, however, the single processor power negated some of the SMP benefits that other vendors had. But I believe that the cost per TPC was not in Sun's favor on these systems. IBM's closest system at the time was probably the S70, but they were a bit later than Starfire. S80 and p690 (regatta) closed the performance gaps.

7. OK, find them. I don't remember any, because for system management, Solaris was, and IMHO still is in the stone ages. I don't believe that Sun ever had decent hardware error handling system, and patch management appears more primitive than AIX. NFS and automount is better, granted, but even for a hardened CLI user, smit is a great fall-back when you can't remember how to do something you touch once in a blue moon. And dynamically loaded and unloaded device drivers allow you to fix a multitude of problems without needing a reboot. And all of the hot-swap hardware makes management easy.

8. Starfire again. How much were they? All IBM power4 systems were paritionable except for the very smallest.

9. Yes? IBMs SMT on Power4 implemented two separately scheduled hardware threads on a single CPU with multiple instruction units, so more like SMT than hyperthreading. The two threads were not completely symmetrical, which is why I said sort-of.

10. Yes. And dynamic LPARing does work. Very well, in fact, as do hot pluggable disks and adapters. Zones is quite different, and I will admit that WPARs were a direct copy of this functionality. I'm not 100% sure, but I am not sure how well Zones split the allocation of CPU and I/O bandwidth between the systems. LPAR overheads, mainly memory (but not CPU), are quite high granted.

11. Starfire again. And again, how much did it cost? And what was available on the smaller systems?

12. Backward compatibility. WTF. I would bet that a 32 bit executable built on AIX 3.1 on RIOS hardware in 1990 has a greater than 70% chance of still running on Power7 running AIX7 twenty years later. How much more backward compatibility do you want? And once you get to AIX 4.3 and AIX 5.1 it will be getting to more than 95% or more. And IBM make it quite clear what features are likely to prevent an application from working. Many software vendors are still compiling their code on AIX 5.1 or AIX 5.2 knowing that their software will work on all later versions of AIX (an example of this is the AIX Toolbox for Linux Compatibility from IBM, that still proclaims to be compiled on AIX 5.1).

I have VERY RARELY (in fact I can hardly remember the last time) had to recompile a program when switching AIX releases unless I wanted to take advantage of new features of a new processor or compiler.

We both obviously have our own perception of what happened and when, but I still believe that the original statement in the article was wrong.

Well

Hi Billl

A few minor errors/things I don't agree with, and that haven't been adressed by others.

6) Jup the Starfire was a great machine. You might say that it was too successful as the SPARC boxes of today is to much like the Starfire, they haven't evolved as fast as perhaps other designs.

I know that many today think that the real nice SMP POWER box was the p690. But the real beauty was the S80/S85 which were a pure breed SMP box'es. The latest versions had 24 Cores and 48 Threads running at the same time so they did get closer to the 64 of the 10K than people remember.

8) The partitioning on the p690 wasn't hardware partitioning, it was Logically partitioning in firmware, it is basically what have evolved into the hypervisor of the POWER7 hardware of today. But the real big jump came with POWER5 and it's hypervisor in 2004, that was quite a jump. And all to many people today still don't understand how to use it properly.

9) The first POWER machine to implement Threading was the RS64-IV which had an implementation of coarse grained multithreading, that was called HMT. Much like later SPARC-64 and Itanium Montecito.

POWER4 didn't have any multithreading, unless you choose to use the SUN term CMT, Chip MultiThreading, which basically means that you have more than one core per chip.

The fist POWER processor to implement SMT was POWER5, in POWER6 multithreading was enhanced allowing the core to issue instructions to two different threads at the same time, in POWER7 a 4 way SMT threading was added.

10) Well didn't get dynamic hardware allocation right ? Sure it's hard to wrestle memory away from a machine that wants to use it :)= But otherwise it IMHO the most dynamic platform when it comes to virtualization. You can actually turn SMT on and of on the fly, and adding and removing virtual processors, adapters etc takes seconds.

Now I think the whole virtualization adds overhead thing is a misunderstanding. If I carve my machine (be it a T5440 or a p690 or a SD using domains Lpars or n/vpars) up into slices I will loose all the idle capacity in all the partitions. When I use virtualization I get access to this, hence often this means that I get access to the 60-80% idle machine time that a Partitioned UNIX machine often has.

11) the G and specially the J series were terrible. I remember once, a place where I worked, we had a Highnode (J50) with 4 cores running as DB servers, the failover machine was a thin node PW2SC 160MHz, in a cluster test one of the UNIX guys forgot to fail back to the high node, next the day the people who used the machine complained that their queries were broken cause they finished to fast. The thin node was just much faster at single threaded workloads.

12) AIX has always been good at compability, only really issue was the 4.3->5.1 64 bit code. But that really was a non issue as nobody really used 64bit executables. Well besides Oracle.. and it was actually more a Oracle issue as they didn't provide a 64bit version of Oracle 8.1.7 that could run on AIX5.1

// jesper