What kind of WAN topology do you have. I.e. how does the HQ logically and physically connect with each branch.
W/o knowing the prior, cannot provide a specific QoS example, but I can present a logical approach.
What you want to do is first, at the HQ side, is shape traffic to each branch to the rate the branch supports (e.g. 6 Mbps, although I suspect some Cisco shapers don't account for L2 overhead, if yours do not, then you have to shape for about 15% less than nominal bandwidth). Then, prioritize that shaped traffic as desired. For starters, just try FQ within class-default.
Ideally you would also want to implement QoS for the HQ's 70 Mbps too, but not all platforms support this. Since 70 Mbps isn't a physical port speed, you would need to shape for that too.
I.e. for HQ you would want a CBWFQ somewhat like this:
policy-map SampleParent
class class-default
shape average 70000000
fair-queue
service-policy SampleChild
policy-map SampleChild
class branch1
shape average 6000000
fair-queue
class branch2
.
.
If you can identify the server(s) doing the large pushes, you can try (again not all platforms support 3 tier) something like:
policy-map SampleChild
class branch1
shape average 6000000
service policy SampleGrandchild
class branch2
.
.
policy-map SampleGrandchild
class PatchServers
bandwidth percent 1
fair-queue
class class-default
bandwidth percent 99
fair-queue
On the branch side, you again need to shape, but assuming branches don't much intercommunicate (if they do, that a very important consideration), something like:
policy-map SampleBranch
class class-default
shape average 6000000
fair-queue
In theory, lots of active branches could over run HQ's bandwidth. The two ways to address that, if it becomes a problem, are either to insure the combined bandwidth for all your branch shapers cannot exceed 70 Mbps (which otherwise limits their to HQ bandwidth, yuk) or have QoS on your WAN's egress to your HQ (usually not possible unless running over a MPLS network).
