Testing… Testing… 1, 2, 3.x... Performance Testing of Pi on NT George Krc Mead Paper

What are we trying to do?

PI 3 Performance Testing n What are we trying to find out? –Is Pi 3 on NT suitable for gauging data applications? –How does DEC Alpha compare with Intel hardware running Windows NT? –How big a box do we need? –How much network bandwidth will be used?

Test Configurations n Configuration #1: “Prioris” –Digital Prioris 6200 MX Server Dual Pentium Pro 200MHz Processors 128mB RAM 10gb Hard Disk Fast-wide SCSI controller PCI 10/100mb Ethernet Carrd

Test Configurations n Configuration #2: “Alpha” –Digital AlphaStation 600 5/266 Single Alpha 266MHz Processor 128mB RAM 10gb Hard Disk Fast-wide SCSI controller PCI 10/100mb Ethernet Carrd

Test Configurations n Configuration #3: “Dell” –Dell PowerEdge 6100 Server Four Pentium Pro 200MHz Processors 256mB RAM 10gb Hard Disk Fast-wide SCSI controller PCI 10/100mb Ethernet Carrd

Test Conditions n Pi Interface on Separate NT API node n Duplicate data sent to test systems n Gauging systems scans every 30 seconds n Data transmission for 15 seconds n Typical data rates in 30 seconds –6633 snapshot events –6236 archive events –26 archive posts

More Test Conditions n Pi version used was 3.1, Build 2.81 n CPU usage for four Pi modules (PISnapSS, PiNetMgr, PiArchss, PiBaseSS) n Statistics gathered with NT Perf. Monitor n Data gathered over 500 seconds intervals n Total CPU usage measured across multiple processors

Test #1 n Compare the CPU Utilization of the Alpha and Prioris Servers under normal data gathering conditions

CPU Utilization Pi Data Gathering

Test #1 Inferences n Prioris outperformed the Alpha by about 2:1 on equal processor speed basis n Testing in production environments can be erratic

Test #2 n Compare the CPU Utilization of the Alpha and Prioris servers and Profile client response time under two conditions: –One Sensor, 250 Scans, 480 Boxes, 30 second wait –Three Sensors, 250 Scans, 480 Boxes, 60 second wait

CPU Utilization With Client Profile Access

Response Time Client Profile Access

Test #2 Inferences n A single client increases Alpha utilization by a factor of 2:1 n Multiple profile continues to increase utilization, but by less than a factor of 2:1 n CPU utilization advantages on the Prioris give better user response times n The Alpha could support 8 continuous profiles, the Prioris about 15

Test #3 n Compare the despooling capabilities of the Alpha and Prioris servers

Despooling Test One Hour of Data (2300 records)

Test #3 Inferences n The Alpha despools significantly faster than the Prioris n The Prioris is limited by I/O throughput

Test #4 n Compare the CPU Utilization of the Dell and Prioris Servers under normal data gathering conditions

CPU Utilization Pi Data Gathering

Test #4 Inferences n The Dell matches the Prioris on a per- processor basis n The Dell maintains a 2:1 performance ratio n The application scales well to a quad processor system n The Dell gives much better performance of the snapshot subsystem (extra memory???)

Test #5 n Compare the CPU Utilization of the Dell and Prioris servers and Profile client response time under two conditions: –One Sensor, 250 Scans, 480 Boxes, 30 second wait –Three Sensors, 250 Scans, 480 Boxes, 60 second wait

CPU Utilization With Client Profile Access

Test #5 Inferences n The Dell demonstrates a better than 2:1 performance increase over the Prioris with twice the processors n The CPU utilization reduction does not translate to a lower response time to the client application n The Dell could support up to 60 continuous profile client accesses

Test #6 n Determine Network Loading under 4 conditions –No Pi Activity –Data Gathering Only –60 Second Client Profile Access –30 Second Client Profile Access

Network Utilization

Test #6 Inferences n Data gathering does not put a significant loading on the network n Client access puts an extremely high loading on the network n 100mb ethernet will be required for the client application

SO WHAT???? –Pi 3 on Windows NT is suitable for gauging data –Both Alpha and Intel are suitable, but the Intel architecture is more cost-effective –The minimum reasonable configuration is a dual Pentium Pro, 200 MHz system with 256 mB RAM and a 100mb ethernet card

AND... –A reasonable limit for the Pi tag count is 10,000 tags –10mb ethernet is insufficient for the application –Pi 3.1 Build 2.81 had a few bugs