Test Day:2009-09-17 Virtualization Hugepages
|Thursday Sep 17, 2009||All day||#fedora-test-day (webchat)|
What to test?
If you come to this page after the test day is completed, your testing is still valuable, and you can use the information on this page to test huge pages support and provide feedback.
John Cooper is your host for today.
The following people have also agreed to be available for testing, workarounds, bug fixes, and general discussion:
- Chris Wright
- add your name here
What's needed to test
- A fully updated Fedora 12 Rawhide machine. See instructions on the main test day page.
- At least one guest image installed before the test day (suggested reading - Virtualization_Quick_Start)
This is the procedure I used to create the initial patch which allows libvirt to recognize/generate a huge page backed guest xml definition. NB: While fairly low-level and useful to unit test, it is however not a mechanism directly visible to a typical user.
The goal here was to allow libvirt to request guest backing by huge pages, which are essentially of 2MB size vs. that of a standard 4KB page. Doing so offers a significant performance benefit in certain application scenarios.
Prepare the Host
Populate the huge page pool of a size suitable to support the guest image(s) which will be created:
# grep Huge /proc/meminfo HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB
# echo 500 > /proc/sys/vm/nr_hugepages
# grep Huge /proc/meminfo HugePages_Total: 500 HugePages_Free: 500 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB
Note the above may take a considerable amount of time on a machine with fragmented physical memory. So it is best to do so as soon after boot as possible. Also on machines with limited memory, populating a smaller number of pages may be necessary.
Having created the free huge page pool, mount hugetlbfs on the host. If the mount point doesn't exist, create it first:
# mkdir /dev/hugepages # mount -t hugetlbfs hugetlbfs /dev/hugepages
Note the mount above must be in place before launcing libvirtd as the daemon currently checks for a hugetlbfs mount only upon startup. So if the daemon is currently running, terminate it:
# kill `cat /usr/local/var/run/libvirtd.pid`
Then (re)start the daemon. Doing so in separate window running in the foreground will allow diagnostics to make their way to stdout:
Launch the Guest
To launch the guest conventionally from virsh:
# <path_to_virsh>/virsh --connect qemu:///system
At this point a guest must be defined by specifying an XML definition (more on this below), for example:
virsh # define /etc/libvirt/qemu/hp_danpb-on.xml Domain foo defined from /etc/libvirt/qemu/hp_danpb-on.xml
In the above example the guest is tagged with the name "foo" in the associated XML definition:
virsh # list --all Id Name State ---------------------------------- - foo shut off
The guest may be launched via:
virsh # start foo Domain foo started
And a VNC connection to the guest console can be made via:
# vncviewer localhost:5900
If all goes well the guest should launch successfully with its image backed by huge pages. [Note it won't unless the guest XML definition specifies huge page usage correctly as below. But proceeding here is instructive in any event.]
Successful launch of a huge page backed guest may be evidenced by observing the huge page free pool decreasing:
# grep Huge /proc/meminfo HugePages_Total: 500 HugePages_Free: 481 HugePages_Rsvd: 247 HugePages_Surp: 0 Hugepagesize: 2048 kB
In the likely case HugePages_Free == HugePages_Total take a look at the XML definition for the guest, For example:
virsh # dumpxml foo <domain type='qemu'> <name>foo</name> <uuid>4c58c2a6-1b52-688e-bcfb-e57159f50961</uuid> <memory>524288</memory> <currentMemory>524288</currentMemory> <vcpu>1</vcpu> <os> <type arch='x86_64' machine='pc'>hvm</type> <boot dev='hd'/> </os> :
The above does not specify a memory backing mechanism and therefore defaults to backing by 4KB pages. To specify huge page backing a <memoryBacking> clause is needed:
virsh # dumpxml foo <domain type='qemu'> <name>foo</name> <uuid>4c58c2a6-1b52-688e-bcfb-e57159f50961</uuid> <memory>524288</memory> <currentMemory>524288</currentMemory> <memoryBacking> <hugepages/> </memoryBacking> <vcpu>1</vcpu> <os> <type arch='x86_64' machine='pc'>hvm</type> <boot dev='hd'/> </os> :
To add this to the XML definition, edit the corresponding file to add the <memoryBacking> clause as above, undefine, and redefine the guest:
virsh # undefine foo Domain foo has been undefined
virsh # define /etc/libvirt/qemu/hp_danpb-on.xml [edited XML def] Domain foo defined from /etc/libvirt/qemu/hp_danpb-on.xml
virsh # start foo Domain foo started
This should result in a huge page backed guest launch which may be verified as above.
There was a modification to the default disposition of selinux genfscon fs types affecting (among others) hugetlbfs in the kernel 2.6.29-2.6.30 timeframe. This manifests as failure of chcon(1) on hugetlbfs files. Correction requires a selinux policy change for hugetlbfs and a corresponding kernel fs change. Neither of which have been conclusively tested as of this writing on prospective FC12. Thus there is a possibility SELINUX may need to be disabled to allow successful launch of a huge page backed guest.
Issues that were identified