VMware Cloud Community

NetIOC / 10GbE musings...

Trying to come up with a good "rule of thumb" for vDS w/ NetIOC on 10GbE environments...  Looking for feedback/etc. on this - based on some thought and research on the web.

Say host has 2 x 10GbE links; no IP storage traffic and no FT:

vDS (NetIOC Shares:  Management = Normal, vMotion = Low, VM Traffic = High)

| ->  Management / HA; load balancing = originating virtual port ID, no failback, active:  vmnic0/vmnic1

| -> VM Network 1; load balancing = LBT, active:  vmnic0/vmnic1

| -> VM Network 2; load balancing = LBT, active:  vmnic0/vmnic1

| -> vMotion - A; load balancing = orig port ID, active:  vmnic0, standby: vmnic1 (egress shaping:  avg bw = 4,000,000 Kbit / peak = 8,000,000 Kbit / burst = 1,500,000 Kbyte)

| -> vMotion - B; load balancing = orig port ID, active:  vmnic1, standby: vmnic0 (egress shaping:  avg bw = 4,000,000 Kbit / peak = 8,000,000 Kbit / burst = 1,500,000 Kybte)


Want management to be stable during network interruption; so no failback after failover.

Want VM Traffic to have the highest priority

Want vMotion to have max platform performance

- 2 vMotion port groups / vmk ports (vMotion-A and vMotion-B)

  - NetIOC shares only apply to ingress, so need to prevent vMotion from saturating links and disrupting VM Traffic.
  - Egress math:  8 max concurrent vMotions w/ 10GbE per host.  We don't want to let the egress of many to saturate one...

     - Let's say 8 is max, on "average" there will be 4 vMotion tasks for planning, and we don't want to ever saturate a single 10GbE link w/ vMotion traffic (so keep the math simple 8 x 1GbE = 8Gb peak for vMotion and 2Gb unallocated for everything else).  Let's say we want to support 30 seconds of peak vMotion activity at a time (burst).  So, average bw is 4,000,000 then peak would be 8,000,000 and the 30 second burst of the difference between average and peak would be 1,500,000KBytes (convert bits to bytes).

Question - is the egress shaping really necessary to prevent the saturation of vMotion traffic from many hosts to one?  Why wouldn't the lower shares of NetIOC on the ingress satisfy that requirement?

Tags (4)
0 Kudos
0 Replies