Tuesday May 11, 2010

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

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