That's surprising. You'd think spreading the workload start over 10 seconds would lower the size of spikes (integrated over a second) by at most a factor of 10.
But the above point is still true: many jobs take a few minutes to run. 60s of dispersion in start time is better than nothing, but you really want more.
(In this case, things are still quantized to a minute boundary, so you'd really want both).
> That's surprising. You'd think spreading the workload start over 10 seconds would lower the size of spikes (integrated over a second) by at most a factor of 10.
If the delay is at the reading side, away from Akami, through a cache, perhaps 10 concurrent requests for X would result in ten lots of data transfer as it isn't in cache yet, but 10 with a short delay is enough to prime the local cache on the first request before the rest start.
There are a number of reasons a sudden glut of activity could balloon bandwidth or CPU/memory costs more than you might expect.
Without a chunk more detail about the system in question, this is just random speculation of course.
But the above point is still true: many jobs take a few minutes to run. 60s of dispersion in start time is better than nothing, but you really want more.
(In this case, things are still quantized to a minute boundary, so you'd really want both).