Ardent Performance Computing

Latest Post

Understanding CPU on AIX Power SMT Systems

This month I worked with a chicagoland company to improve performance for eBusiness Suite on AIX. I’ve worked with databases running on AIX a number of times over the years now. Nevertheless, I got thrown for a loop this week.

TLDR: In the end, it came down to a fundamental change in resource accounting that IBM introduced with the POWER7 processor in 2010. The bottom line is twofold:

1. if SMT is enabled then the meaning of CPU utilization numbers is changed. the CPU utilization numbers for individual processes mean something completely new.
2. oracle database 11.2.0.3 (I haven’t tested newer versions but they might also be affected) is not aware of this change. as a result, all CPU time values captured in AWR reports and extended SQL traces are wrong and misleading if it’s running on AIX/POWER7/SMT. (I haven’t tested CPU time values at other places in the database but they might also be wrong.)

On other unix operating systems (for example Linux with Intel Hyper-Threading), the CPU numbers for an individual process reflect the time that the process spent on the CPU. It’s pretty straightforward: 100% means that the process is spending 100% of its time on the logical CPU (a.k.a. thread – each hardware thread context on a hyper-threaded core appears as a CPU in Linux). However AIX with SMT is different. On AIX, when you look at an individual process, the CPU utilization numbers reflect IBM’s opinion about what percentage of physical capacity is being used.

Why did IBM do this? I think that their overall goal was to help us in system-wide monitoring and capacity planning – however it came at the expense of tuning individual processes. They are trying to address real shortcomings inherent to SMT – but as someone who does a lot of performance optimization, I find that their changes made my life a lot more difficult!

History

Ls Cheng started a conversation in November 2012 on the Oracle-L mailing list about his database on AIX with SMT enabled, where the CPU numbers in the AWR report didn’t even come close to adding up correctly. Jonathan Lewis argued that double-counting was the simplest explanation while Karl Arao made the case for time in the CPU run queue. A final resolution as never posted to the list – but in hindsight it was almost certainly the same problem I’m investigating in this article. It fooled all of us. CPU intensive workloads on AIX/Power7/SMT will frequently mislead performance experts into thinking there is a CPU runqueue problem at the OS level. In fact, after researching for this article I went back and looked at my own final report from a consulting engagement with an AIX/SMT client back in August 2011 and realized that I made this mistake myself!

As far as I’m aware, Marcin Przepiorowski was the first person to really “crack” the case when and he researched and published a detailed explanation back in February 2013 with his article Oracle on AIX – where’s my cpu time?. Marcin was tipped off by Steve Pittman’s detailed explanation published in a December 2012 article Understanding Processor Utilization on Power Systems – AIX. Karl Arao was also researching it back in 2013 and published a lot of information on his tricky cpu aix stuff tiddlywiki page. Finally, Graham Wood was digging into it at the same time and contributed to several conversations amongst oak table members. Just to be clear that I’m not posting any kind of new discovery! :)

However – despite the fact that it’s been in the public for a few years – most people don’t understand just how significant this is, or even understand exactly what the problem is in technical terms. So this is where I think I can make a contribution: by giving a few simple demonstrations of the behavior which Steve, Marcin and Karl have documented.

CPU and Multitasking

I recently spent a few years leading database operations for a cloud/SaaS company. Perhaps one of the most striking aspects of this job was that I had crossed over… from being one of the “young guys” to being one of the “old guys”! I certainly wasn’t the oldest guy at the company but more than half my co-workers were younger than me!

Well my generation might be the last one to remember owning personal computers that didn’t multitask. Ok… I know that I’m still working alongside plenty of folks who learned to program on punch-cards. But at the other end of the spectrum, I think that DOS was already obsolete when many of my younger coworkers starting using technology! Some of you younger devs started with Windows 95. You’ve always had computers that could run two programs in different windows at the same time.

Sometimes you take a little more notice of tech advancements you personally experience and appreciate. I remember it being a big deal when my family got our first computer that could do more than one thing at a time! Multitasking (or time sharing) is not a complicated concept. But it’s important and foundational.

CPU Time on Single CPU Multi Tasking System
  CPU color time for program P1

So obviously (I hope), if there are multiple processes and only a single CPU then the processes will take turns running. There are some nuances around if, when and how the operating system might force a process to get off the CPU but the most important thing to understand is just the timeline pictured above. Because for the rest of this blog post we will be talking about performance and time.

Here is a concrete example of the illustration above: one core in my laptop CPU can copy 13GB of data through memory in about 4-5 seconds:

$ time -p taskset 2 dd if=/dev/zero of=/dev/null bs=64k count=200k
204800+0 records in
204800+0 records out
13421772800 bytes (13 GB) copied, 4.73811 s, 2.8 GB/s
real 4.74
user 0.13
sys 4.54

Continue reading →