Description of problem: - RX and TX queue size are configured on director templates NovaLibvirtRxQueueSize: 1024 NovaLibvirtTxQueueSize: 1024 - this is also correctly set in nova.conf tx_queue_size=1024 rx_queue_size=1024 - however the libvirt domain XML only has RX settings [shtiwari@supportshell virsh]$ grep -e queue -e UUID virsh_-r_dumpxml_instance-00000e74 <driver name='vhost' queues='2' rx_queue_size='1024'/> Instance UUID is: virsh_-r_dominfo_instance-00000e74:UUID: 2f9201e3-df6a-4099-9277-4101db156d10 Version-Release number of selected component (if applicable): RHOSP16.1 Actual results: tx_queue_size is not present in libvirt xml created for guest VM during server creation and hence not propagated to guest VM. only rx_queue_size is propagated. Expected results: tx_queue_size should also be updated and propagated to guest VM with the value configured and present in nova.conf. Additional info:
DFG:Compute will review the requested support exception on October 28th. Since the RFE has been filed for `tuned`, closing this as a duplicate of the tuned RFE. *** This bug has been marked as a duplicate of bug 2015044 ***
Re-opening since we might have to convert this into an RFE for tx_queue_length in OSP 18.
Summary from bug-triage call: - Firstly, libvirt's documentation is very clear that QEMU only supports setting up `rx_queue_size` for interface type "vhostuser". However the user here was using interface type "vhost". So, Nova (and libvirt) ignoring "rx_queue_size" for "vhost" is expected behaviour. - Given the above, *if* there is a valid use-case to have "rx_queue_size" support for interface type "vhost", then the first step here is to have the support for it in QEMU. [TODO] So please file a QEMU RFE needs to be filed first, with a clear mention of the use-case.
(In reply to Kashyap Chamarthy from comment #16) > Summary from bug-triage call: > > > - Firstly, libvirt's documentation is very clear that QEMU only supports > setting up `rx_queue_size` for interface type "vhostuser". However > the user here was using interface type "vhost". So, Nova (and > libvirt) ignoring "rx_queue_size" for "vhost" is expected behaviour. > > - Given the above, *if* there is a valid use-case to have > "rx_queue_size" support for interface type "vhost", then the first > step here is to have the support for it in QEMU. > > [TODO] So please file a QEMU RFE needs to be filed first, with a > clear mention of the use-case. Err, a bad typo. I mean "tx_queue_size" (not "rx_queue_size") in the entire comment above.
(In reply to Kashyap Chamarthy from comment #17) > (In reply to Kashyap Chamarthy from comment #16) > > Summary from bug-triage call: > > > > > > - Firstly, libvirt's documentation is very clear that QEMU only supports > > setting up `rx_queue_size` for interface type "vhostuser". However > > the user here was using interface type "vhost". So, Nova (and > > libvirt) ignoring "rx_queue_size" for "vhost" is expected behaviour. > > > > - Given the above, *if* there is a valid use-case to have > > "rx_queue_size" support for interface type "vhost", then the first > > step here is to have the support for it in QEMU. > > > > [TODO] So please file a QEMU RFE needs to be filed first, with a > > clear mention of the use-case. > > > Err, a bad typo. I mean "tx_queue_size" (not "rx_queue_size") in the > entire comment above. Shravan? Can you please remind us if you've filed the QEMU RFE here?
Let's close this for now pending the QEMU work. If/when that gets completed, we can reopen this RFE, or file a new one to enable the qemu support in Nova.