Comparison with SPECPower Benchmark

Laboratory Home
Test Results
This Site
Email List
Neal Nelson
& Assoc. Home

The following comments identify areas of similarity and areas of difference for the Neal Nelson Power Efficiency Test and the SPECPower Benchmark.


SPECPower Benchmark

Neal Nelson Power Efficiency Test

Power Usage Measured
     At the wall power outlet with an external power monitoring device. At the wall power outlet with an external power monitoring device.
Power Usage Measured at Multiple System Load Levels
     Data collected at 11 levels: Idle plus 10 samples that correspond to 10%, 20%, 30% ... 100% of the server's maximum throughput (as measured by SPECPower). Data collected at 6 levels: Idle plus five fixed transaction arrival rates of 2,400, 4,800, 7,200, 9,600 and 12,000 transactions per minute. The use of fixed transaction arrival rates encourage "apples to apples" comparisons between servers like "when both systems were processing transactions at the rate of 4,800 TPM, system 'A' used 133.0 watts and system 'B' used 145.6 watts."
Operating System Software
     Tests may be run with either Windows or Linux. All tests are run with SUSE Linux Enterprise Server version 10.
System Tunables
     System tunables vary between tests. Tunables are disclosed in the full report. All tests are run with identical tunables.
Disk Input/Output
     No disk I/O during test. Substantial amount of disk I/O dring test.
Application Software
     A single stand alone Java application. Multiple application programs including the Apache2 web server and the MySQL relational database management system plus custom "C" language programs executing as ".cgi" modules.
Number of Client Machines
     One. Thirty two.
Transaction Types
     Inventory transactions. Credit card transactions.
Number of Transactions per Screen/Batch
     SPECPower creates and processes batches of 1,000 transactions at a time. Simulated web clients process each transaction as a single HTML request/response sequence.
Number of Active "Users"
     An small active "user" count that is equal to the number of Java applications configured during a test (normally 1-4). A large active "user" count with 100-500 web clients simultantously processing transactions.
Record Locking
     No record locking is performed. Row level locking on 50% of the transactions processed.
Network Connections
     A small number of static network connections. One connection pair for each copy of the Java application that is running on the server being tested (normally 1-4 copies). These connections are established at the start of the test and remain in place during the test sequence. A large number of dynamic network connections. Each of the 100-500 web clients simultaneously creates a connection with Apache2, transmits a request, waits for a reply and then terminates the connection. This sequence is then immediately repeated for the duration of the test.
Context Switches
     Very few context switchs. The Java application programs accept batches of 1,000 transactions from existing connections and then post them to memory tables. There are few interrupts and very few context switches. Many context switches. Up to 500 users are all communicating via the network, connecting to and disconnecting from MySQL and writing data to the disk drives at the same time there is a very high number of context switchs.
Memory Footprint
     The SPECPower Java application has a relatively small and static memory footprint. It is possible that much of it will fit in the 4 megabyte Intel Core Level 2 memory cache. The Neal Nelson Power Efficiecny Test has a huge and complex memory footprint with Apache2 accepting up to 500 requests, MySQL executing with its multi-gigabyte shared memory segment while the operating system kernel is busy performing all of the network and disk I/O.

Copyright © 2006-2008
Neal Nelson & Associates
Trademarks which may be mentioned on this site are the property of their owners.