For reference, see  bug 638257 .  One of the problems leading
to that failure was, in essence, that results from the DUTs
arrived faster than the ability of gs_offloader to offload and
clear them.

gs_offloader has parameters for allowing more/less parallelism,
so the first thing to look at is whether things can go faster
simply by increasing the number of simultaneous gsutil subprocesses.

I think we decided to attack https://bugs.chromium.org/p/chromium/issues/detail?id=638641 first and maybe do this later. Adding Allen to this one as well. 

It's relatively easy to increase parallelism.  If that gives
us more throughput, we really ought to do it.  Fixing  bug 638641 
only helps us with one specific kind of failure.  There are
other kinds of problems for which improving gs_offloader performance
is our best bet for staving off failure.

Moving to OKR as it doesn't sound urgent. Needs more investigation 

Are we still running into problems here?

I seem to recall some where also related to large volume of crash dumps (eating up upload bandwidth?)