Oracle PeopleSoft Payroll (NA) Sun SPARC Enterprise M4000 and Sun Storage F5100 World Record Performance

The Sun SPARC Enterprise M4000 server combined with Sun FlashFire technology, the Sun Storage F5100 flash array, has produced World Record Performance on PeopleSoft Payroll (North America) 9.0 benchmark.

A Sun SPARC Enterprise M4000 server with four new 2.53GHz SPARC64 VII processors and a Sun Storage F5100 flash array is 33% faster than the HP rx6600 (4 x 1.6GHz Itanium2 processors) on the PeopleSoft Payroll (NA) 9.0 benchmark. The Sun solution used the Oracle 11g database running on Solaris 10.
The Sun SPARC Enterprise M4000 server with four 2.53GHz SPARC64 VII processors and the Sun Storage F5100 flash array is 35% faster than the 2027 MIPs IBM Z990 (6 Z990 Gen1 processors) on the PeopleSoft Payroll (NA) 9.0 benchmark with Oracle 11g database running on Solaris 10. The IBM result used IBM DB2 for Z/OS 8.1 for the database.
The Sun SPARC Enterprise M4000 server with four 2.53GHz SPARC64 VII processors and a Sun Storage F5100 flash array processed 250K employee payroll checks using PeopleSoft Payroll (NA) 9.0 and Oracle 11g running on Solaris 10. Four different execution strategies were run with an average improvement of 25% compared to HP's results run on the rx6600. Sun achieved these results with 8 concurrent jobs using only 25% CPU utilization while HP required 16 concurrent jobs with a 88% CPU utilization.
The Sun SPARC Enterprise M4000 server combined with Sun FlashFire technology processed 8 Sequential Jobs and single run control with a total time of 527.85 mins, an improvement of 20% compared to HPs time of 633.09 mins.
The Sun SPARC Enterprise M4000 server combined with Sun FlashFire technology demonstrated a speedup of 81% going from 1 to 8 streams on the PeopleSoft Payroll (NA) 9.0 benchmark using the Oracle 11g database.
The Sun FlashFire technology dramatically improves IO performance for the PeopleSoft Payroll benchmark with significant performance boost over best optimized FC disks (60+).
The Sun Storage F5100 Flash Array is a high performance high density solid state flash array which provides a read latency of only 0.5 msec which is about 10 times faster than the normal disk latencies 5 msec measured on this benchmark.
Sun estimates that the MIPS rating for a Sun SPARC Enterprise M4000 server is over 2742 MIPS.

Performance Landscape

250K Employees

System	Processor	OS/Database	Time in Minutes			Version
System	Processor	OS/Database	Run 1	Run 2	Run 3	Version
Sun M4000	4x 2.53GHz SPARC64 VII	Solaris/Oracle 11g	79.35	288.47	527.85	9.0
HP rx6600	4x 1.6GHz Itanium2	HP-UX/Oracle 11g	81.17	350.16	633.25	9.0
IBM Z990	6x Gen1 2027 MIPS	Z/OS /DB2	107.34	328.66	544.80	9.0
HP rx6600	4x 1.6GHz Itanium2	HP-UX/Oracle 11g	105.70	369.59	633.09	9.0

Note: IBM benchmark documents show that 6 Gen1 procs is 2027 mips. 13 Gen1 processors were in this config but only 6 were available for testing.

500K Employees

System	Processor	OS/Database	Time in Minutes			Version
System	Processor	OS/Database	Run 1	Run 2	Run 3	Version
HP rx7640	8x 1.6GHz Itanium2	HP-UX/Oracle 11g	133.63	712.72	1665.01	9.0

Results and Configuration Summary

Hardware Configuration:

Software Configuration:

Benchmark Description

The PeopleSoft 9.0 Payroll (North America) benchmark is a performance benchmark established by PeopleSoft to demonstrate system performance for a range of processing volumes in a specific configuration. This information may be used to determine the software, hardware, and network configurations necessary to support processing volumes. This workload represents large batch runs typical of OLTP workloads during a mass update.

To measure five application business process run times for a database representing large organization. The five processes are:

Paysheet Creation: generates payroll data worksheet for employees, consisting of std payroll information for each employee for given pay cycle.
Payroll Calculation: Looks at Paysheets and calculates checks for those employees.
Payroll Confirmation: Takes information generated by Payroll Calculation and updates the employees' balances with the calculated amounts.
Print Advice forms: The process takes the information generated by payroll Calculations and Confirmation and produces an Advice for each employee to report Earnings, Taxes, Deduction, etc.
Create Direct Deposit File: The process takes information generated by above processes and produces an electronic transmittal file use to transfer payroll funds directly into an employee bank a/c.

For the benchmark, we collect at least four data points with different number of job streams (parallel jobs). This batch benchmark allows a maximum of eight job streams to be configured to run in parallel.

Key Points and Best Practices

Why Sun Storage F5100 is a good option for Peoplesoft NA Payroll Application

Accelerate your Payroll Performance with F20 PCIe Card

Disclosure Statement

Oracle PeopleSoft Payroll (NA) 9.0 benchmark, Sun M4000 (4 2.53GHz SPARC64) 79.35 min, IBM Z990 (6 gen1) 107.34 min, HP rx6600 (4 1.6GHz Itanium2) 105.70 min, www.oracle.com/apps_benchmark/html/white-papers-peoplesoft.html Results 10/13/2009.

Comments:

First off, you got the numbers wrong, the F5100 based system is NOT faster than the HP 62-disk HDD system, it is slower; 68.07 minutes for HDD and 79.35 minutes for your Flash Array.

http://www.oracle.com/apps_benchmark/doc/peoplesoft/performance-report/ps9-na-pay-9_ora_hp_rx6600.pdf

Secondly...who the hell are you kidding? You used 40 SLC SSDs -AND- 12 x 15K HDD array...and even then, a conventional 62x15K HDD array was faster than your Flash array?

What the....???? Do you really think people are too stupid to check your numbers?

Posted by Typo or on October 15, 2009 at 11:31 AM PDT #

The 68.07 minutes you are reporting is the total run01 time for the Pay Sheet, Pay Calculation and Pay Confirmation steps. If you add in the Print Advise and Direct Deposit steps HP's total time for run01 is 81.17 minutes. The total time for all 5 steps on Sun was 79.35 minutes, so we are faster. Regarding the rest of your inquiry you can draw your own conclusions regarding the fact that Sun executed run01 with only 8 concurrent streams utilizing 25% cpu while HP executed 16 streams with 88% cpu utilization. Suffice it to say that with this benchmark the M4000 and F5100 Flash Array was not even breaking a sweat.

Posted by Vince Carbone on October 15, 2009 at 12:28 PM PDT #

Vince,

Re:

>> "HP's total time for run01 is 81.17 minutes. The total time for all 5 steps on Sun was 79.35, so we are faster"

Wooowwww...four significant digits of precision here, that's an impressive way to tell us that Sun's configuration of 40 x SSD plus 12 x 15KHDD is 2.293% faster than 58 HDDs without Flash.

Not even close to 33% faster, like it says above, but a measley 2.3% faster. You'll correct the post before HP's lawyers call, right?

And your configuration cost 4x as much, right? Lesseee here, do I need my slide rule to do an ROI analysis?

>> Suffice it to say that with this benchmark the M4000 and F5100 Flash Array was not even breaking a sweat.

Ok...then why did you put 40 SSD's instead of 20 into the F5100?

And...with 40 SSDs vs 58 HDDs, can you explain why your configuration also needed 12-disk array of 15Krpm disks to go with the 40 SSDs? Didn't I just read that each one of your FMOD SSD's was supposed to be good for 17K IOPS? 17K IOPS is about 75 times faster than a 15K HDD -- so what gives? Seems like these SSD's your selling are only about 1.2x faster than HDD.

FYI Vince...have you heard?

"In two process generations, before the end of 2010, NAND Flash will have higher latency than spinning disk"

- Michael Cornwell, Sun's Lead Technologist for Flash, from his August 2009 presentation at the Flash Memory Summit.

Posted by Typo or on October 15, 2009 at 06:00 PM PDT #

[Trackback] I've just read through several comments regarding the F5100. Is it just me or is common sense on vacation at the moment at some locations. I will go through some of this comments and dissect them here. Some voices with criticsmA commentator at Stora...

Posted by c0t0d0s0.org on October 17, 2009 at 04:13 AM PDT #

Typo or -- writes:
> FYI Vince...have you heard?
>
> "In two process generations, before the end of 2010, NAND Flash
> will have higher latency than spinning disk"
>
> - Michael Cornwell, Sun's Lead Technologist for Flash, from his August 2009
> presentation at the Flash Memory Summit.

You took this out of context.

You need to read it in fuller context. This was about Corwell scolding the manufactures on their direction and NOT the value of flash.

http://www.eetindia.co.in/ART_8800581641_1800009_NT_71261a52_2.HTM

"According to Cornwell, there is another problem: Besides cost, NAND vendors are chasing Moore's Law at the expense of the needs of the enterprise market."

Posted by Brad on October 19, 2009 at 05:43 PM PDT #

Brad,

I don't need to read about Cornwell's presentation from the news reports, I was there for his presentation, and I have a copy of it. You should ask him for one.

http://www.flashmemorysummit.com/English/Conference/Proceedings_Chrono.html

From his slides:

===============================

As NAND Manufacturers chase Cost per GB it is coming at a cost:

>Endurance - 1/10th of 3 years ago
>Write Speed - 1/4 performance of 2 years ago
>Read Speed - 1/6 performance of 2 years ago

> NAND will have higher latency than HDD in 2 Generations
> Very few sub-50nm designs capable of supporting Enterprise applications
> Non-endurance failures biggest issue

=================================

The performance degradation Cornwell was talking about is confirmed here, by Toshiba's top Flash engineer, Ken Takeuchi:

http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=4585977

This is why Flash SSD is only capable of reaching a maximum of 2.5x performance of mechanical HDD, as demonstrated in the benchmarks you cite.

From here on out, Flash will get dramatically slower as Flash capacity increases. This is fundamental to NAND Flash device physics and is true for SLC and MLC. By the time we get to the 1xnm lithography node (end of 2010), NAND flash will have higher latency than HDD -- as Mr Cornwell points out.

We will be there by the end of 2010, and even in the most optimistic adoption scenarios, there in not enough "Enterprise SSD" volume in the world to change the silicon foundry economics that force the industry to pursue Moore's law.

Are you guys planning on publishing any detailed cost-benefit analysis on these comparative system results? For example, since the two systems offer identical performance, what is the energy consumption comparison between the 58 HDD system and the 40xSSD + 12xHDD system? How long would it take the much costlier F5100-based system to pay for itself in terms of energy savings?

Finally, why does the blog post above still say "33% faster" when it's only 2% faster?

Posted by Typo or on October 20, 2009 at 10:03 AM PDT #

BestPerf: Performance & Best Practices

Tuesday Oct 13, 2009