XenServer 8 lack of performance

[Robert Versteegen1709152295]

With the implementation of Xenserver for our VDI workload we a strugling with performance of the VDI.

From hour own investigation we think the problem has something to do with the numa assignment.Why: we see at many times that de vcpu assignment is not assigned to a single socket, but assign accros cores on both sockets. This leaves the preffered single socket numa.

[Robert Versteegen1709152295]

(was'nt ready typing)

With the implementation of Xenserver for our VDI workload we a strugling with performance of the VDI.

From hour own investigation we think the problem has something to do with the numa assignment.Why: we see at many times that de vcpu assignment is not assigned to a single socket, but assign accros cores on both sockets. This leaves the preffered single socket numa.

 

Hardware hypervisor

Dell PowerEdge R7525

CPU 2x AMD EPYC 75F3 32-Core Processor

RAM 1024Gib Ram

Storage: SSD 900GiB

HT active resulting in 128 cores

 

Bios is set to maximum performance mode.

Additional settings tested in the bios settings, non of them giving a beter performance

• Numa nodes per socket (default 1)
• L3 cache as numa domain disabled/enabled, tested with enabled, gives a lot of little numa's
• Memory interleave on/off

 

Citrix Intrastructure

Citrix functional level is 2203 LTSR.

 

Configuration of the Xenserver is updated to the latest XenServer 8 release (as of 7/11/2024)

 

Within Xenserver the following configurations have been made

• Read cache on the storage, and Domain0 memory enlarged to 32Gb for more read caching en thus offloading
• Checked and configured setting in xenopsd.conf file with numa-placement=true
• Configured xe host-param-set uuid=<uuid >numa-affinity-policy=best_effort

This changed a littlebit we see it tries to do a numa, although many times it is still across the 2  cpu sockets

 VDI VM configuration

• Windows 10 22h2, all windows updates
• 4vCPU. Tried 4 sockets-1 core/2 sockets - 2 cores/1 socket - 4 cores setup . 4 sockets with 1 core seems to perform best.
• 14GiB memory
• Machines are setup with MCSIO with write caching in RAM 4 GiB, and 10GiB cache disk on the local SSD storage.
• Optimization is done with the Citrix optimizer
• XenTools are installed and up to date, enabled I/O optimization

 

The vm limit is set to 64vm's on this host.  64x4=256 vCPU's. With a 2:1 setting.

If we reduce the number to 4 VM's on this host. The performance is still not well.

We tested with modifying the sched-cred  to  "xl sched-credit -s -t 10 -r 1000"  this helped a bit in performance

Is there some logging options or other thinks we can do to check the lack of performance