[ovs-discuss] Open vSwitch with DPDK: an inquiry about the "dpdk-socket-mem" option

정세연 jsy0906 at postech.ac.kr
Wed Jun 21 07:08:48 UTC 2017


Hi everyone,

I'm trying to construct a NUMA-aware OVS-DPDK testbed with two VMs on each
different NUMA node.

I found this article [1] and knew that for this kind of testbed setup,
there are three types of memory; QEMU mem, OVS mem and DPDK mem (Figure 3.
in the article).

Originally, I simply thought that if I reserve 2048*2M hugepages in the
host and then I start OVS-DPDK with dpdk-socket-mem=1024, 2048, it means
that I can spawn a 1GB VM in the NUMA 0 and two 1GB VMs in the NUMA 1 with
the memory backing option of QEMU.

After reading the article and ovsdb docs [2],
I suspect that the dpdk-socket-mem option may correspond to only
preallocation of the OVS mem, so I have to reduce the numbers in the option
or increase the host's hugepage amount to reserve QEMU memory for VMs.

Please advise me about the relationship with "dpdk-socket-mem" and "OVS /
QEMU memory" in the first article [1].

Thanks in advance,

[1] https://software.intel.com/en-us/articles/vhost-user-
numa-awareness-in-open-vswitch-with-dpdk
[2] http://openvswitch.org/support/dist-docs/ovs-vswitchd.conf.db.5.html

SY Jeong
ᐧ
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openvswitch.org/pipermail/ovs-discuss/attachments/20170621/b180ee10/attachment.html>


More information about the discuss mailing list