The Performance Management of Large Windows NT System

1
The Performance Management of Large Windows NT
System
Des Atkinson Metron Technology Ltd desa_at_metron.co.
uk
2
Contents

• Introduction
• Scalability
• NT and mixed workloads
• Capacity planning of large NT systems
• Management issues
• Conclusions/questions

3
Scalability

• Clusters
• MSCS
• Other offerings
• Symmetric multi-processing
• Large memory systems
• Application scalability
• NT v5.0 enhancements

4
Clusters

• Microsoft Cluster Server
• Shared-nothing, failover model
• 2 non-identical nodes
• Includes application failover
• Scalability of MSCS lies in the future
• Other providers, e.g. Digital Enterprise Cluster

5
Symmetric Multi-Processing

• Standard product supported up to 4-CPU systems
• NT 4.0 Enterprise Edition took this up to 8 CPUs
• More scalable versions available from OEMs
• Asymmetry of NT SMP implementation

6
(No Transcript)
7
Large Memory Systems

• NT is a 32-bit OS
• Supports up to 4GB of virtual memory
• 2 GB for users
• 2 GB reserved for system
• Enterprise Edition increased user share to 3 GB
• NT 5.0 will offer 64-bit VLM

8
Application Scalability

• Exchange Server v5.5
• Removal of 16 GB limit on size of the Message
  Store
• Support for automatic failover in MSCS
• SQL Server v7.0 (Sphinx)
• Full dynamic locking
• Parallel threads
• VLM support
• Support for automatic failover in MSCS

9
NT v5.0 Enhancements

• Active Directory and DFS
• Support for I2O
• Event Trace Facility
• Discussed more fully later
• Microsoft Management Console
• Disk quotas, Kerberos, IPSEC etc.

10
NT and Mixed Workloads

• NT Server not a Mainframe OS
• No GOAL mode as in MVS
• No tools at the OS level to arbitrate between
  workloads
• Plenty of tweaks available at the application
  level

11
Capacity Planning

• NT Performance Monitor
• Shortcomings in PerfMon itself
• Shortcomings in underlying metrics
• NT v5.0 Event Trace
• Capacity planning of clusters

12
NT Performance Monitor

• PerfMon is an application - users can write their
  own
• Metrics have shortcomings
• Log facility is nearest you get to an historic
  performance database
• No built-in planning capability

13
NT 5.0 Event Trace

• Comes with NT 5.0 Beta 2
• Programming tool to allow applications to be
  instrumented
• Will give lots of new metrics
• Overhead very low?
• To be part of WBEM

14
Following Slides based on talk given at US CMG 97

• Authors are Jee Fung Pang and Melur Raghuraman of
  Microsoft

15
Data Requirements for CP

• Process/Thread Termination between sampling
  intervals
• Per Process I/O Counts
• Per Process/Thread per device I/O
• Transaction Boundaries
• File Based Collection
• Per Process/Thread Hard Page Fault Counts

16
NT 5.0 Instrumentation

• Events traced
• Process/Thread Create/Delete
• Physical Disk I/O and Page Fault
• Detailed Metrics can now be derived
• Detail metrics for process, threads and disks
• Allows per device metrics for every thread and
  process
• Bottleneck analysis tools
• e.g. hotfiles, hotdisk, topcpu with drilldown

17
Trace for Applications

• API available
• Allows event tracing to be built into the
  application
• Can provide data such as
• transaction response time, throughput
• transaction-level metrics of CPU and I/O
• Will Microsoft include these in their servers,
  e.g. SQL Server?

18
Capacity Planning of Clustered Systems

• CP of clusters has already been provided on
  OpenVMS and UNIX
• Unit resource usage
• Tracking transactions
• Multi-tier client-server configurations

19
Running the Enterprise on Windows NT

• Account management issues
• Dealing with Microsoft
• Support issues
• What channels are available
• Technical manageability
• Internal support issues
• Compaq/Digital and others
• The new one-stop suppliers?

20
Conclusion

• Until now, Windows NT was not ready to run the
  enterprise
• NT v5.0 should represent a major advance in the
  enterprise
• Wise to wait until SP2 is out unless you feel
  very brave (early 2001?)
• How will the competition respond?