Zones Roadmap Brainstorming Summary March 2, 2007 Solaris 8 Brand Zone Observability Really this is two problems: * Better observability needed inside of zones (i.e. better kstat virtualization) o Better compatibility with agentry o Existing stat tools become more useful * Better observability needed from the global zone o For system deployers and capacity planning o For ISVs who wish to do control plane stuff Memory containment * Improve rcpad o Things are much better following DuckHorn but we may need to do more. o Current implementation can look broken when rcapd gets "stuck" doing work in the kernel + Parallelizing should improve performance + Estimate a few weeks of work o Improve rcapstat/rcapd interaction + Stats don't update often enough * Memory Sets o Need to resync with VM2 team: will VM2 use pools? CPU containment * Unify pools and processor sets o Remove need to choose between pools mode and psets mode * Rewrite poold in C to remove the Java dependency * Consider unification of poold and intrd * libpool and poold - understanding newer processors o We have support for NUMA today, but not CMT * Unify poold, intr, and possibly more into cpumgmtd? Refactor Zones Packaging * Today, zones are in SUNWzoneu and SUNWzoner * Suggestion is to move zones into core packages (SUNWcsu, SUNWcsr) o Various implementation simplifications result in libc o Makes zones available on all systems, not optional o Need to investigate whether this makes sense wrt. IPS project user.* resource controls * Hugely useful for classic time-sharing, build-server sharing * default for "projects on the fly" Enabling Agent-Based Control of Zones * Move configuration into SMF repository * Each zone becomes an SMF instance * Develop JMX based agent in concert with Visual Panels and SMC teams * Discuss agent requirements with 3rd parties as well. * Consider libzonecfg_public.so-- a refactored API which we could support Accounting * "Zone" accounting records would vastly simplify calculating chargeback * Work with community to develop some free and open tools which consume exacct data. (the feedback we got from the field was that Xacct has vanished?) * Consider retiring SYSV accounting o If not, it may need some love Storing zones on nfs servers * Now that we have attach this will be easier * Complexity of kernel work is basically unknown Templates * Provide a set of pre-configured templates for users to choose from * Alter default template to be (more) fork-bomb proof: o 50% physical memory cap o 50% swap limit o 1000 LWPs max o ... * Master template idea: One place to set common resource controls for all of your zones. Attach and patch / Attach and upgrade Zones console improvements * Customers still asking for console logging * Add addition console login features I/O bandwidth containment investigation