Server Performance Issues:Bottlenecks and Enhancements

1
Server Performance IssuesBottlenecks and
Enhancements

• By
• Yemi Ogunrombi
• IS8030 - Integrated Computing Systems

2
Introduction

• The subject of server performance is very broad
  and a major on-going issues in the use of IS/IT
  systems.
• Generally performance issues are detected usually
  when users experience slow response from their
  IT/IS systems. This can be due to one or more
  reasons.

3
Introduction

• Commonly most server systems run either Unix,
  Novell Netware, NT. The focus in this
  presentation is on Unix
• Depending on the infrastructure of the
  organization, there can be one or more servers
  with a number of users connected via varying
  types of network topologies.

4
Monitoring your System

• There are two main strands to understanding a
  system
• The first entails producing a regular collection
  of reports on the machine usage
• The second entails taking a user viewpoint.
• What applications do the users run?
• Could a job be deferred until overnight, or is it
  important for the user to obtain the output
  within the next hour?
• Understanding the business requirement is
  important at this stage

5
Monitoring your system

• In Unix there are 7 major resource types to be
  monitored and tuned for performance
• CPU
• Memory, Virtual Memory
• Disk Sub-system, Interface, Controller
• Communication Lines
• I/O
• Network Time
• Applications Programs

6
First Considerations

• When investigating a possible performance
  problem
• First, determine whether the system is under
  stress or only lightly loaded. (The sar -q,
  uptime and ruptime commands will show averages of
  system load)
• If your system is working hard, you then need to
  identify which processes are the most demanding
  by getting the percentage of the CPU and physical
  memory used by each process (using the top or ps
  commands)

7
First Considerations

• Second, determine how hard the CPU is working
  (iostat will report the percentage of time that
  your CPU is idle)
• If the CPU is lightly loaded, the CPU will spend
  some time being idle.
• If the system is stressed, and the CPU has
  significant idle time, then there is performance
  problem. The CPU may be waiting whilst the system
  is paging, or there may be I/O or network
  bottlenecks.

8
Memory Performance

• A loaded system with significant CPU idle time
  could be caused by memory shortages. (Memory
  usage can be monitored by running vmstat or sar
  -wpgr).
• If there is a high level of swapping and paging,
  then the system is short of memory and is moving
  whole processes to disk
• When a Unix Server is configured, you allocate
  some part of your disk space to be used for
  swapping and Paging

9
Memory Performance

• Possible Solutions
• (The heart of your Unix operating system where
  your operating system configuration is defined is
  the Kernel)
• Remove configuration defined for unused devices
• Configure your system for the maximum number of
  concurrent users that will be using the system.
  The Kernel has internal tables. The size of these
  tables depends on the maximum number of users the
  system can support.

10
Memory Performance

• Possible Solutions
• Minimize buffer cache if this will not adversely
  impact your I/O performance (the buffer cache is
  the area of temporary storage for data moving to
  or from the disks and is used to optimize disk
  access by minimizing transfers)
• Share Paging and Swapping activities evenly
  across the system by distributing swap space over
  as many disks as possible
• Probably the easiest solution is to just buy more
  memory

11
Disk Performance

• Problems with I/O subsystem can severely affect
  system performance
• Memory problems resulting in high paging and
  swapping will increase I/O traffic, therefore it
  is essential to solve memory problems before I/O
  issues.
• Issues might be due to unbalanced disk activity,
  therefore there is need to redistribute your I/O
  activities more evenly among the disks

12
Disk Performance

• Reorganization of filesystems to ensure that
  heavy activities are on the fastest disk drive
  and uses faster controllers
• Large critical files should be places on a file
  system with a large blocksize. Commonly accessed
  files should not be placed in deeply nested
  sub-directories
• Disk fragmentation causes poor I/O performance

13
Disk Performance

• The number of disks that are used should be
  looked into. If a 10GB of disk space is required,
  it is better to have five 2GB disks than one 10GB
  disk.
• With more disks, you have more spindles spinning
  and thus more data can be written to the disks at
  the same time.
• How these multiple disks are arranged is also
  important.
• The RAID technology helps to address some of the
  disk problems such as disk stripping and
  parallelism

14
Tuning the Application

• The way the application code is written is one of
  the primary and common causes of performance
• To detect this, it is important to track which
  jobs are resource intensive and run a trace on
  the process/transaction
• If the process is inevitably resource intensive,
  then if it can be run as a batch during low
  machine usage times

15
Tuning the Application

• Improve or rewrite the code
• This is common in database applications where
  creating an index on a large table could cause
  querying it to change from half an hour to
  seconds

16
CPU Performance

• With the advent of faster processors, typically
  the CPU itself is no longer the problem
• The amount of level 2 cache can have significant
  effects on performance

17
CPU Performance

• However, if the CPU has near 100 utilization,
  but users still have poor response times then the
  following alternatives should be considered
• Reschedule CPU intensive jobs
• Limit the CPU time a process can use with the
  limit command
• Separate the users and application unto separate
  servers
• Upgrade the system

18
Network Performance

• If previous check does not solve the problem, the
  issue could be the network. (use netstat,
  netpmon, nfssat shows network with input and
  output packets and number of collisions)
• It might be necessary to reorganize the network
• There might be slow network components, cards,
  cables etc
• If communications is over national
  telecommunications systems, then that could
  contribute to the performance

19
Conclusion

• Performance issues might be due to one or more
  bottlenecks
• It is important that software applications are
  tested for potential performance issues and tuned
  before implemented.
• RAID technology is advised for disk arrangement
• Thorough planning can help avoid a lot of
  performance issues

20
References

• www.circle4.com/jaqui/papers/webunuk.html
• www.tssp.co.uk/documents/supplements/unixsyssper.h
  tm
• www.cssnw.com/CPUPERF.htm
• www.metron.co.uk/tips/unix.html
• www.networkcomputing.com/shared/server3.html