ࡱ> URdO)\[NZTPicturesC+PowerPoint Document(SummaryInformation(877  !"#$%&'()*+,-./0123456V9:;<=>?@ABCDEFGHIJKLMNOPQRSyWXYZ[\]^_`abcdefghijklmnopqrstuvwxz{|}~ѻϯٿogYw]ww>uy[L$ "pBskZ*Ht*cJ͍8S!}4E5 M98kz-0Wr(΃pDkڂi:k:r좓5.tٳ*_P~0ad ^{[`׹g&qӍp=lgUKXF4 s2ǯQyNZkp-d "zK_|E&ko^\<\Ύ {I:YpgrEN?<~lyau>H=3/>zIc{=zOtU;3϶99ցӐnҏ׫ O-hۧeFnSg%L*Y,vz;^sN0T;+4]Vze%hT?q,&;Џw]&4U#xYO$YZ2Tvą=ȲÂHk =!xTqI ;,t6@!V'! i%vqdqb 휞訊}忪j^qة|}?ٻ\,026gw-_;˞ޓmuyK쫬P׾=ó=`_z, Ć?}!Xx,M;G=ʳ sϲ;+ެ}n3\"yuQEV\|㭠@E}qտ_g, LP>o҇:Z<:ZxéEVn&# Req>Q<ѡx^60EYduyqZX셌.gŽ9\iv025|9Z?ZF~~ !{.]WW>}- ͽ/~}- ޷Yc8 ʯ ү9÷$cۿ]zUo/iRԅSs?2<;FXAE(/ue/JlͲ>Ϗ)?<:׹|Z#qx`F)ʒgI5*c>.u.j',H66>/sW{oBlOQ={0T4()SZ|r644zOiy+Xʧi8>0N[*u@b6媭ZZʧI54a N ĦOb0 .oE*Հ2[N%#1/fwY;oݥv/*Ln3ިUnD]vDv=tF`@Ǻ5;`nb{`?8m5A21^ko0zuzvʙ]yv)hZxNvDnxym7{M|9Sd:go9{^ǯ6S5 %45HW>\pHb݉60vW&1"zo' 5uN }z>+>};TPor9{)4=XolE{ " ٟn Xӎbw8J·2bb2WX3Q2b.7a:> &=81$ G ]퇤gn|ۤP񼲷# ϱ/(ř&3̝( HC4Qu d.g2+2Ju5]X08#h VQn ]Q%ufx25Ʌ0LXȤe%߉s(\+w*8T 1لh ֲp\Mfͅ{M8dz3&. ?|ԡܨ2ě2f Ҝ0Lo<#sPH6id;@gnO1ez۶x>pzDn1-nE;;)>jZ+ؼ~y=ƙhZ1ԊLVgk0`H*rC"@J`*p?5Xy`ceb-A\hm B Њ_e&1)ˤd=S j^R] }A 6%'6& H.2«YI4:|!+Y2!ax YB %.UW\άu5ϗ2'߽v9c-m!X降 R>?RU98!yzfSVf֥x^h܉)\ 6^5DlmfX`vZ , 7 USs9V{i{I)]Z&9[ Ŵ,?)A.ҮCa(5K(n@=te :`^.Rܢ%|s5wm/ZYѴk"k6֩hn+՛8F Q*ڴ]l)xnk;ubݘޟj4}/ z ߊk]I Ԙ0wY~^s ҵm7%(n{mŴ# h0ŽiG찈Qހj^{j7z(6n:Ri~6LۊzE̥euǕHb--P U1t^`[[RLȀI2$sF4wrA&_fy]oμ-ڽQ8{q^\@ž=f+{;pJ|s3;U?RI>\<@c?E`0?`u ?Z~t7q<^V$({,rGB_m2gQ 9Lhf݁f(GddsPvirh0>gMlP =1f~ Lh6Nr`-NWuv^9l :drZ .-3"t&,LXOt_1]6(,#LpaTо{)#Rz.0 hElɋN^ķz!Ce0Hg΄ |BB;3eAx2c-I#9~e7~wm0%q=pX"0q0Y(l'qq nFwX'O|]c>1"1HV}>FK>5 PqZVkǀ>FogKLu]JP(^  h-k,$pt<ɱѭtKc%JxZm[R@1퀁Ce>8[O@XIrƓ˂fq tOg. +YhѦ_a#FBC%ˊi>?oq9BuFmP'p;r}5Î/of\7Q3y4Q] `yi2& z2ABZNpxȡf^HpLqD #&8KTwIэJ,L;'ßNxMX0s*U*dܢR,U2V*k*4JP TKרt7o}4(~): $Zl,+Tٖ.cy.EӋ$ &_ֈVxNCMh~ӲLIDe1&$+ iڠ[HԙKdI4[% ЄNI" 2:I!s>;rpdQpH/D r) P 8Y#@B[St69َZ-U;_{ËarEZrHՇv8!%,?d@P"f,Uc/+'~Rá헣*/g,x+ﾇ,4^ I4q΃mroG/Br7U ,Tɏ?G- /Ύb&nG;(F#lQL;)X%ҵ4%:L<DAƴsco^; S%OM9 \mxNi|4|}0SG;>cۃc/o5bbSld-l;Ŧ pMMqBS-l6F6L$nd^&T[ =sC"%\v#`$?&v#N%'?ɏ yc5>&7&S *OKƬ1b)yD 04^ÄM(`2N`4f [Ai`T_@MVRtzJՈ$.ijcOD-9@!AoajKAqe%);L& )Z[%U D%Ժ=5x ?}N"#.^~ydGmՈecW T3R~',(k Op =(a6M6D$W1ХL'6hCQu crn<׎B >K6)Ef.i8pXl] _0.U'SHlqR&Yv@uc5=Wu},Yݔg@qA_~9wZUu)F kK=`nѲE/Ŏhh)ڊs/XFeMDğQ?JFUYU'm)UL-+[%Qu HD@CG:M";@"dP!^I7{^*Ae_'BH j"* PBPBE!UBE"T,BBabPJxbIH?jQ Q:8(Hi^3pӢOtybG4tQP>冶 C;J5|g$]6?GsT Pr CgFMKd ȝZB'3r+KۼO㌫0Tͳߪ\o܍sdRxU/p ygqyi GihTggs Hvd"=ʺ{]|X/^U|&~ Gox(=l҆94.K]z%גo93ANW%*KS gxA{S.W1^1/uJ<.!^|UvX\Vo*^+4Ux];\Y/צDhA VD?@-!ؔ$MGHD8h$5,Mn]]k_}v󬯪NsnGF~uy iNdAWrGge9X{q9vlq8To}C|5lMO?wPUږJ{#~[Bzg$98sMS nvƚό`3^3`gWlؠ!P Md_t˟5b[#lOj Xv:O_+K :?7Rt',v 9 .Gp3G)᯽*c'S,Ж!h;`NF,<%mvQ]+&jt85Gcl6?%#m'A@=؅d€Axbv1|->!bϲ"PK(b׫bQu^͢>[-,Dʂ2ˁp Fw] _)o x En~tϓXf?ZmlT0"luږ`+jd\hYZ˹X,-=\)V&^$/uB";?ϋOe ~7ʓGWA"m\6z,0u=G<7Z\!7:g0#TN{)hc:?fVIA ^!qU͛3?+IA\r{ép6dBx{w>6H/#1守! oxew [qDntJ!UZV (\*됛f6ςŸ9{k7řB񜇌.>[(6!&J([L#{s\\ͳg&9l姟dw!5瘘3?uUI@DRM\'g38!lE llWH)$Ͼ#{7tRg`fՑ.jkx3+K~Tǂ_#gt;&vӲBC$`['}N0JUJ7@I^' mTF0g}0r@ 8aq G !e ss+LF9ńqTDFz,:uOVY,©"P=q8h;eoS~N`Qb"bddEVhKɀ*_BC~&U2DL:!L-a7@ 1@2Pc Rp[K.-B8d>ydYl2G\`<9FPWX'x-Mm 69f-k% d [Q(P+M)6l,{\aUٰ]J$j"³~g^JȇIV4RЀ"JE4t#\J>>n6 xXX]q!s$RòBGAQN;J~_9O?ɲDZ"G\{p].ɡpQNq'tBE(nƠLJֹ'z>6("q b`7J'[iW^K>sF,UkqH6@;J]wJu;'L!hysgsٍ-g_<}r;>B\o =)z8O"Zµp-\ ע{G&n5rNR\DNµH$\Df ϯ|%Cqw.:-@Y6wy Mla0'WJ{K g]&I;RrsUul5~ e #xėu9L6lPb#,+p"jN+n)50' g1/Q/^uojC}uFtEzoz}LDN8a Sᣆu KkVbiǵJ떽.%ʖT3Z+vgF \ooVNF M9Kwz3𛵛83Uq] _vv<6hJr\wԲiL\lM% -8}4E&@9^Nך\ %`µp=hqS$\ µHDµH$:a>x}1ux6{0u̸( r/0(  0;[0 0 000$([\{b00 000000000  0=] 0 0 0000 2 3 !A0C0E0G0I0c00000000000000000!%),.:;?]}acdeghijklmnopDTimesngbatsenPsd ackDCLB Helvetica Condensed Black DComic Sans MSCondensed BlackB0DZapf DingbatsCondensed Black@DArialingbatsCondensed Black"PDGenevangbatsCondensed Black a .  @n?" dd@  @@`` 4,  !; QQRS{C Q Qk_$$$$B$4]Vze%hC+ `8c? f@@ uʚ;2Nʚ;g4dds{ dQm Pppp@ <4BdBd`gʚ;<4dddd`gʚ;<4dddd`gʚ;ph___PPT2001D<4X? %)SPossible Tools Directions (?) Binder Bind-time (single-node) optimizations Insertion of virtualization functions (ala AutoPilot instrumentation) Anshu? Keith? Linda? Abstract Parallel Machine Maintenance only Abstract Component Machine Helping define vgrids - provide requirements to systems group Scheduling/mapping in vgrid environment 2-phase: Defining the vgrid, mapping w.r.t given vgrid Component  packaging Converting  regular programs into VGrADS form(s) E.g. given FT system infrastructure, insert optimal checkpoints Anirban? Ken?PPPPPfP7PPrPP>  (  7  r UPossible Future Directions:Immediate Future - Evaluate current scheduling techniques and accurateness of performance models Near Future EMAN 1.4/2.0 has support for Python scripting - a good starting point for us to explore construction of a component framework Idea being programs written in Python invoking Grid deployed EMAN components Choosing the right  vgrid for different EMAN components having diverse data and computational needs Mapping EMAN components onto chosen vgrids EMAN COP construction for vgrids, composition of perf models Pre-installed EMAN components & component registration with GIS Component specialization Important Binder issues - late optimization and tuningNmMdmMdtmpVMilestones and Tasks (from SOW) Year 1: APM: Work with virtualization group on design of virtualizing mapper ACM1: Design distributed binder/launcher ACM2: Design workflow application support Year 2: APM2: Prototype virtual task graph clustering mapper ACM3: Prototype building of executables remotely with libraries ACM4: Prototype cluster workflow mapping Year 3: APM3: Prototype hybrid performance modeler for virtualized tasks ACM5: Prototype library installer ACM6: Initial validation of library installer and mapping strategies with applicationt   ZMilestones and Tasks (from SOW) NYear 4: APM4: Initial application validation ACM7: Prototype library installer with component optimization Year 5: APM5: Prototype integration of fault-tolerance into mapper construction APM6: Studies of integration of grid economies into mapper construction ACM8: Prototype domain-specific optimizations for library-based applicationLcc OPW0Lyonl1< 08  .,=>5/2347:-sx,,  e(HH(d&h  ` 33` Sf3f` 33g` f` www3PP` ZXdbmo` \ғq>9y`Ӣ` 3f3ff` 3f3FKf` hk]wwwfܹ` ff>>\`Y{ff` R>&- {p_/̴>?" dU@ <6?" Z2@%  d   << f MMM n?" dZ(@   @@``PR      ` p>>   (     `txaxa1 ?O%  T Click to edit Master title style! !@  Ztxaxa1 ?  RClick to edit Master text styles Second level Third level Fourth level Fifth level!     SR  s *<61?QyR  s *<61?`~  HA1?"0B  s *޽h ? lb  Blank UM`(  )  Z jxaxa1 ? m  ;Body Text Second Level Third Level Fourth Level Fifth Level     <   Z jUU1?o   Page * Z  ###55FFp  01 ?  B  s *޽h ? a( p 0(     Z jUU1?o  n Page * Z ###55FFB  s *޽h ? a(   bZ (  l  C ЦiO%   l  C i}    <i8c?( f(Chuck s slide)( 2 H  0޽h ? lb   rj0(  r  S iO%   r  S tz    <j8c?  j"(Anirban s Slide)( 2 H  0޽h ? lb   @(  l  C PjO%   l  C jq    <j8c?k AAPM = Abstract Parallel Machine; ACM = Abstract Component Machine B( 2B BH  0޽h ? lb   P(  l  C jO%   l  C j     <P j8c?k AAPM = Abstract Parallel Machine; ACM = Abstract Component Machine B( 2B BH  0޽h ? lbcxp^RЀ3ÿ lHbP  @AL G@;b `B&V  (  7  rtW f>"9IUPossible Future Directions:Immediate Future - Evaluate current scheduling techniques and accurateness of performance models Near Future EMAN 1.4/2.0 has support for Python scripting - a good starting point for us to explore construction of a component framework Idea being programs written in Python invoking Grid deployed EMAN components Choosing the right  vgrid for different EMAN components having diverse data and computational needs Mapping EMAN components onto chosen vgrids EMAN COP construction for vgrids, composition of perf models Pre-installed EMAN components & component registration with GIS Component specialization Important Binder issues - late optimization and tuningNmMdmMdtmpVMilestones and Tasks (from SOW)Year 1: APM: Work with virtualization group on design of virtualizing mapper ACM1: Design distributed binder/launcher ACM2: Design workflow application support Year 2: APM2: Prototype virtual task graph clustering mapper ACM3: Prototype building of executables remotely with libraries ACM4: Prototype cluster workflow mapping Year 3: APM3: Prototype hybrid performance modeler for virtualized tasks ACM5: Prototype library installer ACM6: Initial validation of library installer and mapping strategies with applicationt  ,8+ZMilestones and Tasks (from SOW)NYear 4: APM4: Initial application validation ACM7: Prototype library installer with component optimization Year 5: APM5: Prototype integration of fault-tolerance into mapper construction APM6: Studies of integration of grid economies into mapper construction ACM8: Prototype domain-specific optimizations for library-based applicationLcc,AZPTLyonl1< 08  .,=>5/2347:-sx,,  e(HH(d&h     (  l  C ЦiO%   l  C i   H  0޽h ? lb   0$(  r  S iO%   r  S tz  H  0޽h ? lbr:RZJ\Sj:N^Հ ( r/0(  0;[0 0 000$([\{b00 000000000  0=] 0 0 0000 2 3 !A0C0E0G0I0c00000000000000000!%),.:;?]}acdeghijklmnopDTimesngbatsXenPsd ackDCLB Helvetica Condensed Black DComic Sans MSCondensed BlackB0DZapf DingbatsCondensed Black@DArialingbatsCondensed Black"PDGenevangbatsCondensed Black a .  @n?" dd@  @@`` 4,  !; QQRS{C Q Qk_$$$$B$4]Vze%hC+ `8c? f@@ uʚ;2Nʚ;g4dds{dQm Pppp@ <4BdBd`gʚ;<4dddd`gʚ;<4dddd`gʚ;ph___PPT2001D<4X? %SPossible Tools DirectionsfBinder Bind-time (single-node) optimizations Insertion of virtualization functions (ala AutoPilot instrumentation) Abstract Parallel Machine Maintenance only Abstract Component Machine Helping define vgrids - provide requirements to systems group Scheduling/mapping in vgrid environment 2-phase: Defining the vgrid, mapping w.r.t given vgrid Component  packaging Converting  regular programs into VGrADS form(s) E.g. given FT system infrastructure, insert optimal checkpoints Performance model construction Continue to refine the node performance model strategyPlPPPPfP7PPrPP7Pl>  (  7  r7tW f>"9UPossible Future Directions:Immediate Future - Evaluate current scheduling techniques and accurateness of performance models Near Future EMAN 1.4/2.0 has support for Python scripting - a good starting point for us to explore construction of a component framework Idea being programs written in Python invoking Grid deployed EMAN components Choosing the right  vgrid for different EMAN components having diverse data and computational needs Mapping EMAN components onto chosen vgrids EMAN COP construction for vgrids, composition of perf models Pre-installed EMAN components & component registration with GIS Component specialization Important Binder issues - late optimization and tuningNmMdmMdtmpVMilestones and Tasks (from SOW)Year 1: APM: Work with virtualization group on design of virtualizing mapper ACM1: Design distributed binder/launcher ACM2: Design workflow application support Year 2: APM2: Prototype virtual task graph clustering mapper ACM3: Prototype building of executables remotely with libraries ACM4: Prototype cluster workflow mapping Year 3: APM3: Prototype hybrid performance modeler for virtualized tasks ACM5: Prototype library installer ACM6: Initial validation of library installer and mapping strategies with applicationt  ,8+ZMilestones and Tasks (from SOW)NYear 4: APM4: Initial application validation ACM7: Prototype library installer with component optimization Year 5: APM5: Prototype integration of fault-tolerance into mapper construction APM6: Studies of integration of grid economies into mapper construction ACM8: Prototype domain-specific optimizations for library-based applicationLcc,AZPTLyonl1< 08  .,=>5/2347:-sx,,  e(HH(d&h     (  l  C ЦiO%   l  C i=  H  0޽h ? lbr^.Sn^&Հa+( r/0(  0;[0 0 000$([\{b00 0What We ProposedWhere Are We? Milestones and Tasks (from SOW) Milestones and Tasks (from SOW)Key Issue for This WorkshopPossible Future DirectionsPossible Tools DirectionsLyon  Fonts UsedDesign Template Slide Titles Custom Shows 8@ _PID_HLINKS'A&_m,Sarah Gonzales00000000  0=] 0 0 0000 2 3 !A0C0E0G0I0c00000000000000000!%),.:;?]}acdeghijklmnopDTimesngbatsXenPsd ackDCLB Helvetica Condensed Black DComic Sans MSCondensed BlackB0DZapf DingbatsCondensed Black@DArialingbatsCondensed Black"PDGenevangbatsCondensed Black a .  @n?" dd@  @@`` TL&   !; QQRS{C Q Qk   _$$$$B$4]Vze%hC+ `8c? f@@ uʚ;2Nʚ;g4dds{dQm Pppp@ <4BdBd`gʚ;<4dddd`gʚ;<4dddd`gʚ;ph___PPT2001D<4X? %[)Abstract Machine Layer Research in VGrADS#&Where do we stand? Where should we go?\What We ProposedAbstract Parallel Machine Runs MPI programs as if on a single homogeneous cluster Goals scaled back to near zero in revised statement of work Abstract Component Machine Vision: Application integrates components preinstalled on resources across the Grid Binder processes application remotely on each resource Sophisticated workflow scheduling Schedule entire workflow at once APM scheduling becomes a node in workflow graph High level interface Matlab? Python? S? Application Studies PtPPPLP7P"PQPPPPPtL  7  "Q#] Where Are WeUnder GrADS We have a prototype Abstract Parallel Machine No more work under revised SOW We have developed a prototype Abstract Component Machine Workflow scheduling Includes installed software (thanks, Rich) New distributed binder No high level interface Scheduling graph entered manually Prototype performance model construction We have a prototype implementation of one target application EMAN 3D image analysis application .9+/"*=# .9  +  /"*=#,?`VMilestones and Tasks (from SOW)Year 1: APM: Work with virtualization group on design of virtualizing mapper ACM1: Design distributed binder/launcher ACM2: Design workflow application support Year 2: APM2: Prototype virtual task graph clustering mapper ACM3: Prototype building of executables remotely with libraries ACM4: Prototype cluster workflow mapping Year 3: APM3: Prototype hybrid performance modeler for virtualized tasks ACM5: Prototype library installer ACM6: Initial validation of library installer and mapping strategies with applicationt  ,8+ZMilestones and Tasks (from SOW)NYear 4: APM4: Initial application validation ACM7: Prototype library installer with component optimization Year 5: APM5: Prototype integration of fault-tolerance into mapper construction APM6: Studies of integration of grid economies into mapper construction ACM8: Prototype domain-specific optimizations for library-based applicationLcc,AZ^Key Issue for This Workshop2Where to Draw the Line between Absract Machine Layer and Virtual Grids? Without this, we really cannot proceed Proposal: Virtual grid mechanism defines a collection of resources fixed in time, but distinguished by Type (e.g. machine class, installed software, installed databases) Size (cluster of x nodes, memory per node) Performance (virtual mhz, virtual bandwidth) Vgrid mechanism handles all rescheduling and fault tolerance Abstract machine layer does scheduling for virtual configuration and provides information (e.g. perf model) about app to vgrid layertH' ]H' ]t3UPossible Future Directions:Immediate Future - Evaluate current scheduling techniques and accurateness of performance models Near Future EMAN 1.4/2.0 has support for Python scripting - a good starting point for us to explore construction of a component framework Idea being programs written in Python invoking Grid deployed EMAN components Choosing the right  vgrid for different EMAN components having diverse data and computational needs Mapping EMAN components onto chosen vgrids EMAN COP construction for vgrids, composition of perf models Pre-installed EMAN components & component registration with GIS Component specialization Important Binder issues - late optimization and tuningNmMdmMdtmpSPossible Tools DirectionsfBinder Bind-time (single-node) optimizations Insertion of virtualization functions (ala AutoPilot instrumentation) Abstract Parallel Machine Maintenance only Abstract Component Machine Helping define vgrids - provide requirements to systems group Scheduling/mapping in vgrid environment 2-phase: Defining the vgrid, mapping w.r.t given vgrid Component  packaging Converting  regular programs into VGrADS form(s) E.g. given FT system infrastructure, insert optimal checkpoints Performance model construction Continue to refine the node performance model strategyPlPPPPfP7PPrPP7Pl>  (  7  r7tW f>"9PTLyonl1< 08  .,=>5/2347:-sx,,  e(HH(d&h   (  l  C tr.p  l  C `qr `    H  0޽h ? lb   (  l  C вqO%   l  C pi  H  0޽h ? lb   ( t l  C iO%   l  C pu,  H  0޽h ? lb   ( ( l  C 0-E O%   l  C -E   H  0޽h ? lbrb@ˬîS>Հb+( r/0(  0;[0 0 000$([\{b00 000000000  0=] 0 0 0000 2 3 !A0C0E0G0I0c00000000000000000!%),.:;?]%http://www.hipersoft.rice.edu/grads/}acdeghijklmnopDTimesngbatsXenPsd ackDCLB Helvetica Condensed Black DComic Sans MSCondensed BlackB0DZapf DingbatsCondensed Black@DArialingbatsCondensed Black"PDGenevangbatsCondensed Black a .  @n?" dd@  @@`` TL&   !; QQRS{C Q Qk   _$$$$B$4]Vze%hC+ `8c? f@@ uʚ;2Nʚ;g4dds{dQm Pppp@ <4BdBd`gʚ;<4dddd`gʚ;<4dddd`gʚ;ph___PPT2001D<4X? %[)Abstract Machine Layer Research in VGrADS#&Where do we stand? Where should we go?\What We ProposedAbstract Parallel Machine Runs MPI programs as if on a single homogeneous cluster Goals scaled back to near zero in revised statement of work Abstract Component Machine Vision: Application integrates components preinstalled on resources across the Grid Binder processes application remotely on each resource Sophisticated workflow scheduling Schedule entire workflow at once APM scheduling becomes a node in workflow graph High level interface Matlab? Python? S? Application Studies PtPPPLP7P"PQPPPPPtL  7  "Q#] Where Are We?Under GrADS We have a prototype Abstract Parallel Machine No more work under revised SOW We have developed a prototype Abstract Component Machine Workflow scheduling Includes installed software (thanks, Rich) New distributed binder No high level interface Scheduling graph entered manually Prototype performance model construction We have a prototype implementation of one target application EMAN 3D image analysis application .9+/"*=# .9  +  /"*=#,?`VMilestones and Tasks (from SOW)Year 1: APM: Work with virtualization group on design of virtualizing mapper ACM1: Design distributed binder/launcher ACM2: Design workflow application support Year 2: APM2: Prototype virtual task graph clustering mapper ACM3: Prototype building of executables remotely with libraries ACM4: Prototype cluster workflow mapping Year 3: APM3: Prototype hybrid performance modeler for virtualized tasks ACM5: Prototype library installer ACM6: Initial validation of library installer and mapping strategies with applicationt  ,8+ZMilestones and Tasks (from SOW)NYear 4: APM4: Initial application validation ACM7: Prototype library installer with component optimization Year 5: APM5: Prototype integration of fault-tolerance into mapper construction APM6: Studies of integration of grid economies into mapper construction ACM8: Prototype domain-specific optimizations for library-based applicationLcc,AZ^Key Issue for This Workshop2Where to Draw the Line between Absract Machine Layer and Virtual Grids? Without this, we really cannot proceed Proposal: Virtual grid mechanism defines a collection of resources fixed in time, but distinguished by Type (e.g. machine class, installed software, installed databases) Size (cluster of x nodes, memory per node) Performance (virtual mhz, virtual bandwidth) Vgrid mechanism handles all rescheduling and fault tolerance Abstract machine layer does scheduling for virtual configuration and provides information (e.g. perf model) about app to vgrid layertH' ]H' ]t3UPossible Future Directions:Immediate Future - Evaluate current scheduling techniques and accurateness of performance models Near Future EMAN 1.4/2.0 has support for Python scripting - a good starting point for us to explore construction of a component framework Idea being programs written in Python invoking Grid deployed EMAN components Choosing the right  vgrid for different EMAN components having diverse data and computational needs Mapping EMAN components onto chosen vgrids EMAN COP construction for vgrids, composition of perf models Pre-installed EMAN components & component registration with GIS Component specialization Important Binder issues - late optimization and tuningNmMdmMdtmpSPossible Tools DirectionsfBinder Bind-time (single-node) optimizations Insertion of virtualization functions (ala AutoPilot instrumentation) Abstract Parallel Machine Maintenance only Abstract Component Machine Helping define vgrids - provide requirements to systems group Scheduling/mapping in vgrid environment 2-phase: Defining the vgrid, mapping w.r.t given vgrid Component  packaging Converting  regular programs into VGrADS form(s) E.g. given FT system infrastructure, insert optimal checkpoints Performance model construction Continue to refine the node performance model strategyPlPPPPfP7PPrPP7Pl>  (  7  r7tW f>"9PTLyonl1< 08  .,=>5/2347:-sx,,  e(HH(d&h    ( t l  C iO%   l  C pu,  H  0޽h ? lbr]^ϴUՀ