Your SlideShare is downloading. ×
Preliminary Experiment of Disaster Recovery based on Interconnect-transparent Migration
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×

Saving this for later?

Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime - even offline.

Text the download link to your phone

Standard text messaging rates apply

Preliminary Experiment of Disaster Recovery based on Interconnect-transparent Migration

305
views

Published on

Poster at PRAGMA24

Poster at PRAGMA24

Published in: Technology

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
305
On Slideshare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
2
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. Preliminary*Experiment*of*Disaster*Recovery** based*on*Interconnect9transparent*Migra<on Ryousei(Takano,(Hidemoto(Nakada,(Takahiro(Hirofuchi,(Yoshio(Tanaka,(and(Tomohiro(Kudoh( Na9onal(Ins9tute(of(Advanced(Industrial(Science(and(Technology((AIST),(Japan Summary* Incremental*block*migra<on*!  Quick$recovery$is$desirable$from$catastrophic$events,$e.g.,$ !  An$incremental$block$migraDon$is$a$storage$migraDon$ the$great$eastRJapan$earthquake$on$March$11,$2011.$ mechanism$that$enables$us$to$migrate$a$VM$without$shared$!  We$propose$a$novel$disaster$recovery$soluDon$based$on$ storage.$$A$base$(master)$image$is$shared$between$source$ virtual$machine$(VM)$migraDon$technologies,$combining$ and$desDnaDon,$and$only$the$delta$image$is$transferred$to$ interconnectRtransparent$migraDon$(Ninja$migraDon)$and$ the$desDnaDon.$ incremental$block$migraDon.$ !  We$adapt$it$for$disaster$recovery$on$WAN$environments.$$!  To$demonstrate$the$feasibility,$we$have$conducted$an$ The$base$image$is$stored$in$shared$storage$of$each$data$ experiment$of$VM$migraDon$between$an$Infiniband$cluster$ center,$and$it$is$synchronized$between$data$centers$in$ and$an$Ethernet$cluster$on$an$emulated$WAN$environment. advance.$ !  The$cost$of$storage$migraDon$can$be$significantly$reduced$ Survey*of*IT*infrastructure*damages* since$compute$nodes$of$a$cluster$is$homogeneous.$ Iwate Prefectural Figure 1. Location of the Data$center$1 Synchroniza9on( Data$center$2 University institutions that provided IT in(advance infrastructure damage reports: cmp00 cmp01 cmp04 cmp05 Iwate Prefectural University, Tohoku University, High mem. mem. Energy Acceleration Research Organization (KEK), University VM00 VM01 of Tsukuba, and AIST. The red Tohoku Univ. dot represents the epicenter. base$ Wide$area$ base$ The map is a courtesy of image network image KEK wikipedia. Red regions mem. mem. Tsukuba Univ. experienced JMA seismic str00 str01 intensity of 6 or more, and 4 to VM02 VM03 AIST 6 in yellow and orange regions. cmp02 cmp03 cmp06 cmp07!  Data$collected$about$IT$infrastructure$damages$on$the$great$ Ninja(migra9on(+(incremental(block(migra9on east$Japan$earth$quake$is$summarized$in$the$table$below$[1].$ Figure 3. A disaster recovery scheme using incremental block migration!  IT$equipment$suffered$minimal$damages.$!  30$to$60$minutes$available$for$VM$migraDon$acDviDes. Evalua<on* Distance from Seismic IT equipment. Electrical Power Network Connectivity the Epicenter Intensity damages Iwate Prefectural University 220 km 6- none Power uninterrupted (generators) Redundant links kept connectivity alive !  We$demonstrate$migraDon$of$4$VMs$from$an$Infiniband$ Tohoku 150 km 6- to 6+ none UPS supplied tens of Lost after 28 minutes, cluster$to$an$Ethernet$cluster$by$using$the$proposed$soluDon.$ University minutes due to SINET shutdown !  We$used$two$5$nodeRclusters$connected$with$Gigabit$ UPS supplied tens of KEK 310 km 6- none minutes Data not available Ethernet$via$a$hardware$network$testbed$GtrcNETR1,$which$ Univ. of Tsukuba 310 km 6- none UPS supplied tens of minutes Lost immediately emulates$the$roundRtrip$latency$varied$from$0$to$200$ms$and$ AIST 310 km 6- minimal UPS supplied 15 to 60 minutes Available for 60 minutes the$bandwidth$varied$from$100$Mbps$to$1$Gbps.$ !  The$overhead$of$Ninja$migraDon$is$about$3$seconds.$ !  We$esDmate$that$the$number$of$VMs$which$can$evacuate$in$ Ninja*migra<on* one$hour$are$from$56$(JP$to$US)$to$480$(JP$to$JP).!  Ninja$migraDon$[2]$is$an$interconnectRtransparent$migraDon$ 300$ Host*machine*environment mechanism$to$migrate$a$cluster$of$VMs$between$data$centers$ 250$ 229$ 246$ 251$ 253$ OS Debian$Linux$7.0 equipped$with$heterogeneous$interconnect$devices.$ QEMU 1.4 Elapsed*<me*(second) 200$!  Ninja$migraDon$has$no$performance$overhead$during$normal$ VM*environment 150$ VCPU 8 operaDons,$and$MPI$processes$can$migrate$between$data$ Memory 512$MB centers$without$restarDng$the$processes.$ 100$ Disk 5$GB!  Ninja$migraDon$expands$the$opportunity$and$the$availability$ 50$ 30$ 30$ 33$ 35$ OS ScienDfic$Linux$6.3 for$VM$migraDon$between$data$centers.$ 0$ MPI$lib. Open$MPI$1.6.4 1$Gbps$ 100$Mbps$ OFED 3.5 0$ms$ 10$ms$ 100$ms$ 200$ms$ VM VM VM MPI app. MPI app. MPI app. Figure 4. Experimental setting and result Infiniband node MPI lib. MPI lib. MPI lib. IB TCP .. . IB TCP .. . IB TCP .. . Future*work* IB$HCA virDo_net$ detach reDaEach (Ethernet) VMM VMM VMM VMM !  An$MPI$free$implementaDon$is$a$future$work.$ Migration to Migration to !  We$also$plan$to$evaluate$the$feasibility$of$the$proposed$ Ethernet node Infiniband node disaster$recovery$soluDon$on$a$real$distributed$environment$ VM MPI app. like$PRAGMA$Cloud. Ethernet node MPI(library(switches( MPI lib. the(transport(protocol( [1]$M.$Tsugawa,$R.$Figueiredo,$J.$Fortes,$T.$Hirofuchi,$H.$Nakada,$and$R.$Takano,$“On$the$Use$of$ .. in(coopera9on(with(VM( IB TCP migra9on.( . VirtualizaDon$Technologies$to$Support$Uniterrupted$IT$Services,”$RIDO$2012.$ [2]$R.$Takano,$H.$Nakada,$T.$Hirofuchi,$Y.$Tanaka,$and$T.$Kudoh,$“Ninja$MigraDon:$An$InterconnectR$ transparent$MigraDon$for$Heterogeneous$Data$Centers,”$HPGC$2013$(to$be$published).$ VMM VMM VMM VMM $ Acknowledgement:$This$work$was$partly$supported$by$JSPS$KAKENHI$Grant$Number$24700040$and$ Time 1) 2) 3) 4) JRRAPID$of$Japan$Science$and$Technology$Agency. Figure 2. An overview of Ninja migration. It works with cooperation between a VMM and an MPI runtime system on the guest OSs.