Ibm aix wlm idea
Download as PPT, PDF0 likes1,108 views
This document discusses adding artificial intelligence capabilities to workload managers like IBM's AIX Work Load Manager (WLM) to help address system performance problems. It proposes using monitoring data and fuzzy logic rules to detect issues, identify problematic processes, and dynamically reschedule processes to prioritize important services. Existing system instrumentation and soft computing tools could be integrated with Perl to implement this. However, these ideas are theoretical and soft computing approaches are not widely known or accepted. The goal is to give workload managers more "brains" to autonomously address performance problems based on gathered data and expert knowledge encoded as fuzzy rules.
Ibm aix wlm idea
- 1. IBMIBM WLMWLM unortodoxunortodox ideaidea How to add some brains andHow to add some brains and self tunning into performance controllself tunning into performance controll Damir DelijaDamir Delija 20062006
- 2. What are modern UNIX schedulersWhat are modern UNIX schedulers Based on the old IBM workload manager ideasBased on the old IBM workload manager ideas Separate system resuorces into well defined classesSeparate system resuorces into well defined classes Fine control and statisticsFine control and statistics Dynamic controlDynamic control AIX Work Load Manager (WLM) our target, butAIX Work Load Manager (WLM) our target, but Solaris and others also can be usedSolaris and others also can be used Clusters are also nice targetClusters are also nice target
- 3. What is the problem ?What is the problem ? Basic problemBasic problem No UNIX expertise, because it is not UNIXNo UNIX expertise, because it is not UNIX way of thinkingway of thinking ResultResult Modern schedulers are very sophisticatedModern schedulers are very sophisticated tools but usually not very much used,tools but usually not very much used, because - people don’t think into appropriatebecause - people don’t think into appropriate wayway So there is a problem ….So there is a problem ….
- 4. What can be a solution ?What can be a solution ? Can we add some brains into WLMs behaviour?Can we add some brains into WLMs behaviour? How we can do that ?How we can do that ? How one react to performance problem ?How one react to performance problem ? - find the hog or problem,find the hog or problem, - cure (temporary) problem,cure (temporary) problem, - do data collecticng,do data collecticng, - learn the cause,learn the cause, - remedy the problem permanentlyremedy the problem permanently
- 5. How WLM worksHow WLM works Each process has priority and belongs toEach process has priority and belongs to classclass Each class has set of resources with wellEach class has set of resources with well defined limits (memory, IO, cpu)defined limits (memory, IO, cpu) Process can be dinamycally reassinged toProcess can be dinamycally reassinged to classclass Tight utilisation monitoring by WLM – howTight utilisation monitoring by WLM – how limits are appliedlimits are applied Class limits can’t be brokenClass limits can’t be broken
- 6. So what can we do whenSo what can we do when performance problem arraise?performance problem arraise? ImportantImportant this is a short term firefight solutionthis is a short term firefight solution idea is to get minimal impact on crticalidea is to get minimal impact on crtical servicesservices don’t be democratic let VIP to survivedon’t be democratic let VIP to survive Hog is detected, we can put it intoHog is detected, we can put it into unprivileged class, response andunprivileged class, response and resources are conserved!resources are conserved!
- 7. IdeaIdea 1) Monitor the performace1) Monitor the performace WLM monitoring, RMC, nmon,WLM monitoring, RMC, nmon, Classic UNIX tools (sar, vmstat,iostat ..)Classic UNIX tools (sar, vmstat,iostat ..) Application/system specific toolsApplication/system specific tools 2) Detect problem – data from monitoring2) Detect problem – data from monitoring Human like rules (process P is top CPU consumer for a longHuman like rules (process P is top CPU consumer for a long time) - perfect for fuzzytime) - perfect for fuzzy Crisp rules (Crisp rules (iow >= 45%)iow >= 45%) 3) Correct the problem3) Correct the problem Reschedule/disable process – put into unprivilegd classReschedule/disable process – put into unprivilegd class
- 8. How we can implement rules?How we can implement rules? Fuzzy logicFuzzy logic unfortunate name, but works and it is not resource-expensive tounfortunate name, but works and it is not resource-expensive to implementimplement Rules can be easily presentedRules can be easily presented if A is big and A often then slow down Aif A is big and A often then slow down A This is what we use when we are talking about theThis is what we use when we are talking about the problem !!problem !! Expert knowlwdge about system can be easily build intoExpert knowlwdge about system can be easily build into rules!!rules!!
- 9. Are there tools for such system ?Are there tools for such system ? System instrumentation already existsSystem instrumentation already exists Various soft computing tools in easy toVarious soft computing tools in easy to use languagesuse languages Perl - perfect intergation languagePerl - perfect intergation language Some prototypes done in ksh and tclSome prototypes done in ksh and tcl Java and Phyton also usefullJava and Phyton also usefull Distributed WLM also existsDistributed WLM also exists
- 10. Why it is not yet done ?Why it is not yet done ? It is not a way how things are done :(It is not a way how things are done :( There are some theorethical papers butThere are some theorethical papers but related to SLA and huge complex toolsrelated to SLA and huge complex tools Soft computing is not well knownSoft computing is not well known expecially fuzzy logic methodsexpecially fuzzy logic methods Have you ever tried to talk about “fuzzyHave you ever tried to talk about “fuzzy methods” with upper management ?methods” with upper management ?
Editor's Notes
- #2: >For Simon, but tobe used later with same paper