VMware Fault Tolerance

FT replikuje iba input vstupy klavesnice, mysi, resp. interrupts. Tie su nahravane a na druhej strane znova zrealizovane voci Secondary VM.
Just feed inputs v rovnakom case ako boli vstupy na primary site.
Sekvencne citanie vyraba najviac latencie. Da sa riesit spustanim aktivneho Disk Read pre Secondary VM na cieli. Default je diskove citanie na Secondary VM vypnute, pretoze znamena zdvojenie citacich operacii na poli.
Pri FT sa nepouziva Binarny Preklad! Pre FT musi byt pouzivana Hardware Assisted Virtualization od AMD, resp. Intel.
Pozor spomalenie secondary VM na sekundarnom ESX znamena spomalenie primarnej nody FT.
Ked nemoze posielat primarna VM, FT sa zastavi.
FT CPU overhead je zanedbatelny! :)

Uroven sietovej prevadzky medzi primarnym a sekundarnym ESX pri neaktivnych beziacich virtualizovanych OS:
RHEL (1000Hz) - 1,43 Mbit/sec FT logging traffic
SLES (250Hz) - 0,68 Mbit/sec FT logging traffic
Win 2003 (82Hz) - 0,15 Mbit/sec FT logging traffic

99% CPU/Memory intensive benchmark SPEC
RHEL 1,5 Mbit/sec FT logging traffic
Idle 1,43 Mbit/sec FT logging traffic

Kernel compile 100% 3 Mbit/sec FT logging traffic

Network
64 K packet a 8K message size
Tx FT 60 Mbit/sec FT logging traffic, only acks
Rx FT 990 Mbit/sec FT logging traffic
depends on direction

Netperf

Filebench
lot file reads, huge FT
2k 155 Mbit/sec FT logging traffic,
2kw 3,6 Mbit/sec FT logging traffic
64k 1400 Mbit/sec FT logging traffic
64kw 2 Mbit/sec FT logging traffic

FT zvysuje disk latenciu zanedbatelne - 3,4 mikro seconds!!! :)))

Enterprise Applications:

Oracle 11g - 1CPU dnes urobi na novom hardware viac ako dva stare CPU. Vykon noveho CPU je dvojnasobny. 80 users Swingbench FT log 12Mbit, latency nebadatelna

MS-SQL 2005 - reading from CPU, velmi casto - FT 18Mbits - Response time 65msec na 100ms, throughput bez vyraznej degradacie

Exchange 2007 - latency response - 55ms sa zvysi na 70ms

VMmark - multitier benchmark - 5 VMs, dalsich 5VMs more and more, z inymi workloads. Mail server Exchange, MS-SQL, 65 VMs on one ESX, 20 VMs pod FT. 10,11 FT VMs na jednom ESX je optimalne maximum. 4 VMs max on one ESX recommended, pretoze nieje dobre koncentrovat vela kritickych VMs na jednom ESX.

Roundtrip between ESX servers requirement 1ms. Generalne pre LAN je 1ms maximum roudtrip inac je problem na sieti LAN.

Podpora pre SMP - v najblizsom case urcite nie. Pri SMP sa musia riesit i pristupy do pamate, co je velmi narocne. Pracuje sa na tom, vyvija sa, ale zatial to nevyzera realne na najblizsiu dobu.

Zdroj: http://www.vmworld.com/docs/DOC-3484

http://www.vmware.com/files/pdf/perf-vsphere-fault_tolerance.pdf

http://www.vmware.com/files/pdf/resources/ft_virtualization_wp.pdf