| *** rbudden has joined #scientific-wg | 00:00 | |
| masber | good afternoon, is this the right place to ask about an issue about pinning instances to Numa nodes on openstack? | 02:42 |
|---|---|---|
| *** rbudden has quit IRC | 03:30 | |
| *** priteau has joined #scientific-wg | 08:22 | |
| *** priteau has quit IRC | 11:09 | |
| *** priteau has joined #scientific-wg | 11:10 | |
| *** priteau has quit IRC | 11:11 | |
| *** rbudden has joined #scientific-wg | 12:27 | |
| *** jmlowe_ has joined #scientific-wg | 12:38 | |
| *** jmlowe has quit IRC | 12:40 | |
| jmlowe_ | masber: afaik the resident expert in that kind of thing is Blair | 15:12 |
| masber | I see | 15:13 |
| jmlowe_ | masber: CERN did some trials, I don't think they got much out of numa settings relative to cpu pinning | 15:13 |
| masber | I got my problem sorted | 15:13 |
| masber | my problem was that I enabled this on my images --property hw_cpu_policy=dedicated --property hw_cpu_thread_policy=isolate | 15:14 |
| masber | which "disabled" hyperthreading reducing the number of cores I could use | 15:14 |
| masber | jmlowe_, do you have experience in provider networks? | 15:15 |
| jmlowe_ | a little | 15:15 |
| jmlowe_ | probably as much as anybody else | 15:15 |
| jmlowe_ | what do you have in mind? | 15:15 |
| masber | I was thinking to start playing with provider networks so I can reduce the latency of my network | 15:16 |
| masber | this openstack environment will be living next to an HPC environment so we are ok with all vms sharing same network | 15:17 |
| jmlowe_ | linuxbridge or ovs? | 15:18 |
| masber | we currently use ovs | 15:18 |
| masber | I mean do I still need ovs/linuxbridge with provider networks? | 15:18 |
| jmlowe_ | ok, I'm not nearly as well versed in ovs | 15:18 |
| jmlowe_ | you need one or the other to wire up your instances | 15:19 |
| jmlowe_ | I guess you could use sriov and skip those | 15:19 |
| masber | jmlowe_, ok, I use kolla-ansible which has an option to enable provider networks but I am not sure how it works... | 15:20 |
| masber | so I guess the ovs/llinux bridge will act as a virtual switch which is connected to the physical interface | 15:21 |
| jmlowe_ | as I under stand them, provider networks wire up external vlans into your instances | 15:21 |
| masber | can u have just a flat network? | 15:22 |
| jmlowe_ | ovs would create a virtual switch that has your hpc environment's vlans and your instance's tap interface | 15:22 |
| masber | ok, and how are the vlans setup on the instances? do I need to setup an external DHCP server? | 15:24 |
| masber | or can I use neutron? | 15:24 |
| jmlowe_ | I can tell you about a thing I did, I use linuxbridge so ymmv | 15:25 |
| masber | yes please | 15:27 |
| masber | my problem is that I don't have a test environment and my hardware running openstack is currently used for a couple of projects so I can't test | 15:27 |
| masber | so I need some guidance from here to make my mind around | 15:28 |
| jmlowe_ | We have a 10PB lustre filesystem reexported over nfs, I created a vlan and a RFC <whatever> non routable subnet, I have that vlan tagged on all my compute switch ports, I created the tagged interface on the compute nodes, added physical_interface_mappings for the tagged interface i.e. tg-cie160051:bond0.364 | 15:28 |
| jmlowe_ | I created a network in neutron and used the mapping as they physical_network argument so that it knows to attach that interface on the computes to the switch it builds when adding a interface on that network to an instance | 15:30 |
| jmlowe_ | I set the subnet to match the subnet I created in my datacenter networking gear, used dhcp, no gateway, static route to datacenter router when talking to NFS server | 15:31 |
| jmlowe_ | make sense? | 15:32 |
| masber | I think yes, if I am not mistaken tagged vlan all the way through from your luster to the switch and compute nodes, the neutron knows how to tag the nic of the instance assigned to that network and the DHCP is external | 15:33 |
| masber | could you assign the VLAN through DHCP? | 15:34 |
| jmlowe_ | dhcp is internal, forgot to mention, I also have the mapping on my network nodes so they can attach and serve dhcp | 15:37 |
| jmlowe_ | but that's a minor site specific thing for us, dhcp is local to clusters and doesn't cross those boundaries | 15:38 |
| masber | ok, I think I need time to play with this and digest | 15:39 |
| masber | I am not an expert in openstack networking, just use it | 15:40 |
| jmlowe_ | same | 15:43 |
| bollig | rbudden: jmlowe_: https://aws.amazon.com/blogs/aws/amazon-lightsail-update-launch-and-manage-windows-virtual-private-servers/ . Pretty cheap for a month of service for Server 2012 R2 | 18:08 |
| jmlowe_ | not terrible, little light on the memory in my opinion | 18:10 |
| bollig | windows has always been great at paging memory ;-) | 18:10 |
| *** priteau has joined #scientific-wg | 20:45 | |
| *** flanders_ has joined #scientific-wg | 21:00 | |
| *** priteau has quit IRC | 21:22 | |
| *** priteau has joined #scientific-wg | 21:23 | |
| *** priteau_ has joined #scientific-wg | 21:26 | |
| *** priteau has quit IRC | 21:27 | |
| *** priteau_ has quit IRC | 22:47 | |
| *** flanders_ has quit IRC | 23:19 | |
Generated by irclog2html.py 2.15.3 by Marius Gedminas - find it at mg.pov.lt!