Hardware Reference
In-Depth Information
File%System%
User%Quotas%
Life%Time%
Target%Usage%
5&GB&&
150K&inodes&
Permanent user storage;
automatically backed up.
&
$HOME&
Project&
400&GB&
&3M&inodes&
Large allocated storage; not
backed up.
&
$WORK&
Project&
Large temporary storage; not
backed up, purged periodically.
$SCRATCH&
none&
14&Days&
FIGURE7.1:LustrelesystemquotasandtargetusageonTACC'sStampede
system.
containingabnormallyhighbad-sectorcounts.Aftercompletingtheindivid-
ualdiskburn-intests,thesecondstepoftheprocesswastobuildandtestthe
productionRAIDsetsthatwouldultimatelybeformattedtosupportLustre.
Thesemultipledevice(MD)RAID-leveltestsperformedlargeblockI/Oto
eachofthesixRAIDdevicesperOSSinparallelineight-hourintervals.For
thisparticularhardwareandRAID6softwareRAIDconguration,theperfor-
manceforasatisfactoryMDdevicewas595MB/s.AnyRAIDsetwhichdid
notmeetthisperformancethresholdwasre-examinedtoidentifyslowdrives
inhibitingperformance.Notethatthisapproachfollowssimilardeployment
strategiesonpreviouslarge-scalesystemswhichhaveprovedusefulinisolating
outliersanddetectinginfantmortality[4].Theentireprocesswasrepeatedun-
tiltheRAIDsetperformancewasconsistentacrossallOSSs.Todemonstrate
theresultingperformanceconsistency,Figure7.2presentsthemeasuredMD
writespeedsobtainedfortheOSSsserving HOMEandWORK. Based on
FIGURE 7.2: Average software RAID performance for Lustre OST devices
measured during disk burn-in procedure.
