AIX > Trends > What's New

My Hands-On Experience With POWER8


SMT8 Support and Performance

To get SMT8 support on POWER8, the LPAR needs to run AIX v7.1 TL3 SP3, otherwise the LPAR will run in SMT4 mode. I decided to use nmem, ncpu and my own test program written in C to compare the three LPARs, which all had the same entitlement, VPs and memory. On the S814, one LPAR ran AIX v7 with SMT8 and a second ran AIX v6 with SMT4. The third ran AIX v7 with SMT4 on a 750B (which has 32 x 3.3ghz cores). Memory on the 750 was 1066mhz and on the S822L and S814 was 1600mhz, which puts the 750 memory at about 66 percent of the speed of the S814 memory. Comparing rPerfs for three cores on each shows the 750 LPAR to be estimated at 60.42 percent of the rPerf of the SMT8 AIXv7 POWER8 LPAR and 64.66 percent of the SMT4 AIXv6 POWER8 LPAR. The AIX v6 SMT4 POWER8 LPAR is estimated at 93.44 percent of the AIX v7 SMT8 POWER8 LPAR. So how did they stack up?

Using nmem I saw the following as an overall average (OPS/S K is thousands of operations per second):

		OPS/S K
b814aix1	29820.78		P8 AIX v7 SMT8
b814aix2	27483.75		P8 AIX v6 SMT4
b850nl1	        21708.83		P7 AIX v7 SMT4

Comparatives	OPS/S K
nl1/aix1	73.89%
aix1/nl1	135.54%
nl1/aix2	81.94%
aix2/nl1	122.69%
aix2/aix1	90.43%
aix1/aix2	110.78%

Comparing the OPS/S K, we can see the 750 LPAR performed better than predicted by the difference in mhz in the memory. As we all know, mhz is not everything; variables like buffering, cache sizes and even instruction type all impact how quickly things move through memory. With that said, the memory performance in operations per second was still 22 percent to 35 percent better on POWER8 than it was on POWER7, which is a significant improvement.

Comparatives	rPerf		Results (128 processes)	C Program
nl1/aix1	60.42%		67.95%			59.43%
aix1/nl1	165.5%		147.17%		        168.28%
nl1/aix2	64.66%		74.19%			71.31%
aix2/nl1	154.65%	        134.78%		        140.23%
aix2/aix1	93.44%		91.58%			83.33%
aix1/aix2	107.02%	        109.19%		        120%

When comparing rPerf, I’d expect the 750 LPAR to be between 60 and 65 percent of the POWER8 LPARs and the AIXv7 POWER8 LPAR to be around 7 to 8 percent faster than the AIX v6 POWER8 LPAR. Multiple tests were run with ncpu using varying numbers of processes. The results reported are averages from multiple runs using 128 processes. In all three LPARs, this exceeded the number of threads available to the LPAR and it stressed the CPU accordingly. When comparing the user time from ncpu for 128 processes on each LPAR to the other LPARs, it’s clear that the two POWER8 LPARs seem to scale very closely to how the rPerf scales. The POWER7 LPAR did better than expected using 47 percent more user time than the AIX v7 P8 LPAR instead of 65.5 percent more and 34.78 percent more user time than the AIX v6 P8 LPAR instead of 54.65 percent. Similar results were seen using the C program written to drive CPU.

My Conclusions

I should note that these were very limited tests and there are many more tests left for me to run in the benchmarking suites that I am testing. Additionally, they don’t test all of the functions of the server and are not a true mixed OLTP workload, which is what most run. However, they do provide some initial data that shows that POWER8 appears to scale as expected, both due to improvements in the memory performance as well as CPU performance. Also the jump from SMT4 to SMT8 provides around a 9 percent boost, which is on a par with what’s predicted in the published rPerf. Further, more detailed tests are planned using several test suites for memory, cpu, I/O and network performance. Although preliminary and limited, they provide a window into the performance potential of the new POWER8 scale-out servers.

Overall, the POWER8 experience has been very positive so far. Firmware and HMC updates went smoothly and the server appears to be performing as expected. This provides a level of confidence that you can move from POWER7 to POWER8 while reducing the VPs or cores in LPARs using rPerf comparisons as an approximate scaling factor.

Jaqui Lynch is an independent consultant, focusing on enterprise architecture, performance and delivery on Power Systems with AIX and Linux.



Like what you just read? To receive technical tips and articles directly in your inbox twice per month, sign up for the EXTRA e-newsletter here.


comments powered by Disqus

Advertisement

Advertisement

2019 Solutions Edition

A Comprehensive Online Buyer's Guide to Solutions, Services and Education.

AIX > TRENDS > WHAT'S NEW

POWER5 & APV: Perfect Together

IBM Systems Magazine Subscribe Box Read Now Link Subscribe Now Link iPad App Google Play Store
IBMi News Sign Up Today! Past News Letters