Per-core Performance Myth Busting

IBM continually touts "performance per core" in sales and marketing messages.  IBM implies that higher performance per core will somehow deliver better customer experience.

Oracle's "Optimized System Performance"  vs. IBM's "Per-core Performance Focus"

Customers care about system performance and the ROI of their solution.  Does better "per-core performance" predict better system performance or price/performance?  No, is the simple answer.

Modern server & CPU designers can make various trade-offs on complexity, performance and number of threads & cores.  The best way to address these trade-offs is to look at the integrated system design.

Below are two examples where better "system design" is far more important than a focus on "per-core" performance:

  • Oracle's Sun SPARC Enterprise M9000 server delivered a single-system TPC-H 3000GB world record.
    • beats IBM's Power 595 performance by 20%
    • beats IBM's Power 595 price/performance (3 year TCO: hardware, software, maintenance, etc.)
    • Oracle database load time(4hr 45min) was over 2 times faster than IBM (10hr 2min)!
    For TPC-H, IBM used half the number of cores, but could not deliver better customer value.
  • Oracle's 12-node Sun SPARC Enterprise T5440 server cluster delivered a TPC-C world record.
    • beats IBM's Power 595 (5GHz) with IBM DB2 9.5 database performance by 26%
    • beats IBM's IBM's Power 595 price/performance (3 year TCO: hardware, software, maintenance, etc.) by 16%
    • in addition with the Oracle solution one also has better response time, Oracle's New Order response time was 7.3x faster than IBM!
    For TPC-C, IBM used one-sixth the number of cores, but could not deliver better customer value.

In conclusion, Better ROI is achieved with Oracle's Integrated System design.


Required Disclosure statements:

Sun SPARC Enterprise M9000 (32 procs, 128 cores, 128 threads) 188,229.9 QphH@3000GB, $20.19/QphH@3000GB, database load time 4:43:25, avail 04/10/10.  IBM Power 595 (32 procs, 64 cores, 128 threads) QphH@3000GB, 156,537.3 QphH@3000GB, $20.60/QphH@3000GB, database load time 10:02:25, avail 11/24/09. TPC-H, QphH, $/QphH tm of Transaction Processing Performance Council (TPC). More info www.tpc.org
http://blogs.sun.com/BestPerf/tags/tpc-h

12-node Sun SPARC Enterprise T5440 Cluster (12 nodes, 48 procs, 384 cores, 3072 threads) with Oracle 11g Enterprise Edition with Real Application Clusters and Partitioning, 7,646,486.7 tpmC, $2.36/tpmC, response time new order average 0.168, Available 3/19/10. IBM Power 595 (5GHz Power6, 32 chips, 64 cores, 128 threads) with IBM DB2 9.5, 6,085,166 tpmC, $2.81/tpmC, response time new order average 1.22, available 12/10/08. TPC Benchmark C, tpmC, and TPC-C are trademarks of the Transaction Performance Processing Council (TPC), source: www.tpc.org, results as of 11/5/09.
http://blogs.sun.com/BestPerf/tags/tpc-c

Comments:

How about System Performance comparision between Oracle and IBM :

SAP R/3 SD 2-Tier (SAP R/3 (Sales & Distribution) Standard Application Benchmark)

Product Date Processor Clock (GHz) Chips Cores Users SAPS Steps/Hr Version Status

IBM - Power 780 7-Apr-10 POWER7 3.80 8 64 37,000 202,180 12,131,000 ERP 6.0U (2009) Current

Sun - SPARC Enterprise M9000 18-Nov-09 SPARC64 VII 2.88 64 256 32,000 175,600 10,536,000 ERP 6.0U (2009) Current

Posted by guest on May 11, 2010 at 03:21 PM PDT #

You forgot to add the pricetags.... This article isn't about performance, it's about price/performance.

Posted by Stefan Hinker on May 11, 2010 at 05:51 PM PDT #

SAP is licensed on per user basis, so the bottom line is how much does the system cost, how much performance can it deliver, and how much will it cost over 3 years?
-3YR TCO, price/performance is what matters

-The SAP SD 2-Tier benchmark unfortunately does not have a full hardware configuration disclosure report like TPC-C or TPC-H and therefore we do not know how the IBM Power 780 was configured (besides the basics like CPUs/Cores/RAM) and how much that configuration costs-especially over a 3 year period.

- SAP SD 2-Tier benchmark also doesn't disclose the price of the configuration so we have no idea exactly how and what IBM configured and how much it costs. Theres no full disclosure report to prove if the IBM configuration is customer configurable(off the shelf) or a special IBM "benchmark special" config. 
-Even IBM has stated (http://www-03.ibm.com/systems/power/hardware/780/perfdata.html)┬áthe following with respect to its POWER7 performance results: “The benchmarks and values shown herein were derived using particular, well configured, development-level computer systems. Actual system performance may vary and is dependent upon many factors including system hardware configuration and software design and configuration.”  In other words, the devil is in the details.  Unfortunately, details are not available for SAP benchmark configurations.

- IBM ran DB2 instead of OracleDB -The question is-what is the performance of IBM's Power7 systems with Oracle? There are no Oracle benchmarks to substantiate.

- Does a customer just run SAP?
- If not, where are the high end DB related benchmarks on Power7 and why is there none? What is IBM hiding?  No TPC-H, no 8-socket TPC-C, no SPECJapp Server, no Siebel, etc.

-SAP benchmark isn't I/O intensive and so does not represent how a DB-tier workload would perform on platform.  
SAP benchmark is very CPU bound/dependent- its CPU cache dependent and memory I/O dependent and so clearly high GHZ, 32MB L3 cache and DDR3 memory has a positive impact on performance. The database size of SAP benchmark is small enough to be cached on the Power7 system.  

Finally, the IBM Power System 780 SAP configuration used 8 processors / 64 cores / 256 threads & 1024GB RAM @ 99% CPU utilization while the Sun M9000, 64 processors / 256 cores / 512 threads, 1152 GB main memory ran at 94% utilization so on a per thread basis, IBM's system offers roughly 2x better per thread performance but required 7x more RAM per CPU. Note that the Sun system still has ~5% of CPU utilization/overhead left.

Posted by Phil on May 11, 2010 at 06:12 PM PDT #

Thanks Dave for starting this blog. A per core performance does not necessarily
mean better experience. A casual look at the TPC_H benchmark reveals the value
of SPARC/Solaris/Oracle database integration, esp with the database load time metric.
Oracle integrated systems loads the database in less than 5 hours compared to IBM loads it in more than 10 hours. A simple comparison to prove the value of integration that saves customer time to configure systems, OS and load a database for their DSS.
Hours become even important if you pay someone to get that job done and they bill you twice as much - a more of integrity than integration point though.

Posted by Sandeep Bhalerao on May 13, 2010 at 12:20 AM PDT #

Post a Comment:
Comments are closed for this entry.
About

BestPerf is the source of Oracle performance expertise. In this blog, Oracle's Strategic Applications Engineering group explores Oracle's performance results and shares best practices learned from working on Enterprise-wide Applications.

Index Pages
Search

Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today