120 likes | 240 Views
Scalability Overview. Revised Dec 27 2005. Scalability – What Is It. Predictable performance over a range of loads Predictable does not equal good – just predictable No growth preferred but linear response is acceptable Just knowing performance hit for a load may be enough.
E N D
Scalability Overview Revised Dec 27 2005
Scalability – What Is It • Predictable performance over a range of loads • Predictable does not equal good – just predictable • No growth preferred but linear response is acceptable • Just knowing performance hit for a load may be enough
Scalability – We Document Guidelines • Guidelines are not a guarantee • Tests often ignore network (run on GB unconstrained) • Config, Bugs can confound any guidelines
Scalability - Network • Network factors can slow any client server process down any amount desired • Scalability guidelines assume no network constraints • Factors: • End-to-end round trip time (satellite, excessive hops) • Bandwidth used up • Excessive errors • Config
Scalability – Network – What to look for • Wide swings in CLI ping time • OS reported network load • Trace route
Scalability – CPU Contention • Too much work running on one box • Task queuing, often with CPU close to 100% • Wide unexplained swings in response time
Scalability – Memory Contention • Too much work running on one box • Excessive paging • Commit charge – time close to peak • Wide unexplained swings in response time
Scalability – Resource “x” Contention • Too much work running on one box • Wide swings in response time, often unexplained
Scalability – “max” Guidelines • Not a guarantee that you will like the performance • We test to determine breakpoints – max often is highest count that does not break • We report key performance values at max – you decide • See tests below max value • Guidelines are backed off max, but your “knee of curve” will vary
Scalability • general – total objects*state changes/messages per second per/object • Deglitch chatter – do this as far down the foodchain as possible – define policy at top of foodchain, run the same policy everywhere to the extent possible – use attributes to guide policy
Scalability • WorldView – number of maps, map complexity • CORE – total objects, state changes per second • Event – messages per second, # rules, % forwarded • DSM – total objects, polls per second • Agents – objects monitored, system internal poll rate, internal event rate • uDSM – machines per domain, replication to enterprise, reporting
Scalability • DSM UAM – what items at enterprise level, number of domains, engines, sectors – network • DSM USD – admin and operational views for enterprise and local servers – network! • DSM USD – package size, number of packages, number of targets for a package, delivery windows • SD – Tickets/unit time; updates/unit time; # analysts; # clients connected at once