A SERVICE OF

logo

6
VMWARE WHITE PAPER
To ensure that the secondary virtual machine runs as fast as the primary, it is recommended that:
• ThehostsintheFTclusterarehomogenous,withsimilarCPUmake,model,andfrequency.TheCPUfrequencydifferenceshould
not exceed 400 MHz.
• Boththeprimaryandsecondaryhostsusethesamepowermanagementpolicy.
• CPUreservationissettofullforcaseswherethesecondaryhostcouldbeoverloaded.TheCPUreservationsettingontheprimary
applies to the secondary as well, so setting full CPU reservation ensures that the secondary gets CPU cycles even when there is
CPU contention.
2.4. I/O Latencies
All incoming network packets to the primary, and all disk reads at the primary, are immediately sent to the secondary. However, as
explained in section 1.4, network transmits and disk writes at the primary are held until the secondary acknowledges that all events
that precede the packet transmit or disk write. As a result, the round-trip network latency between the primary and the secondary
aectstheI/Olatencyofdiskwritesandnetworktransmitoperations.SincetheroundtriplatencyinaLANenvironmentisusuallyin
the order of a few hundred microseconds, and disk I/O latencies are usually on the order of a few milliseconds, this delay does
not impact disk write operations. One may, however, notice delays in network ping responses if the response time is shown in
microseconds. For best performance, it is recommended that the round-trip network latency between the primary and secondary
host be less than 1 millisecond.
2.5. Network Link
Since the primary and secondary virtual machines proceed in vLockstep, the network link between the primary and the
secondary host plays an important role in performance. A Gigabit link is required to avoid congestion. In addition, higher bandwidth
network interfaces generally have lower transmission latency. If the network is congested and the primary host is not able to send
trac to the secondary (i.e. when the TCP window is full), then the primary virtual machine will make little or no forward progress. If
the network connection between the primary and secondary hosts goes down, either the current primary or the current secondary
virtual machine will take over, and the other virtual machine will die.
2.6. NIC Assignments for Logging Traffic
FT generates two types of network trac:
• Migrationtraffictocreatethesecondaryvirtualmachine
• FTloggingtraffic
MigrationtrachappensovertheNICdesignatedforVMotionanditcausesnetworkbandwidthusagetospikeforashorttime.
SeparateanddedicatedNICsarerecommendedforFTloggingtracandVMotiontrac,especiallywhenmultipleFTvirtual
machinesresideonthesamehost.SharingthesameNICforbothFTloggingandVMotioncanaecttheperformanceofFTvirtual
machines whenever a secondary is created for another FT pair or a VMotion operation is performed for any other reason.
VMwarevSwitchnetworkingallowsyoutosendVMotionandFTtractoseparateNICswhilealsousingthemasredundantlinksfor
NICfailover.SeeKB article 1011966 for more information.
Adding multiple uplinks to the virtual switch does not automatically result in distribution of FT logging trac. If there are multiple FT
pairs, then trac could be distributed with IP-hash based load balancing policy, and by spreading the secondary virtual machines to
dierent hosts.
2.7. Virtual Machine Placement
FT logging trac is asymmetric: the bulk of the trac ow happens from the primary to the secondary hosts and the secondary host
only sends back acknowledgements. If multiple primary virtual machines are co-located on the same host, they could all compete for
thesamenetworkbandwidthontheloggingNIC.Idlevirtualmachinesconsumelessbandwidth,butI/O-intensivevirtualmachines
can consume a lot of network bandwidth. It can be helpful to place the primary of one FT pair and the secondary of another FT pair
onthesamehosttobalancethetracontheFTloggingNIC.