Following up on te network discussion during the ops team
meeting, I think that the devil is in the details:
* There is a difference between core JANET capacity, MAN
capacity, and MAN<->site capacity. My overall impression
is that overall JANET has capacity, but WAN-JANET and
site-JANET links sometimes are fairly busy, usually
with non grid traffic.
* Looking at long term averages for checking whether a
link is busy is not helpful, because:
- network traffic quantity can be rather bursty;
- latency matters too, when transfers happen;
- specific direction (in or out) matters too.
As to the latter, there is a difference between "major" VOs,
some of which in theory have a steady streams of jobs and
transfers, and other users. But even the major VOs sometimes
send large batches of jobs to T2s that then pull down or push
back specific data sets and how fast that happens matters too.
So I'd think that one should look at percent of capacity over a
number of fairly short periods (e.g. days), and independently
for incoming and outgoing traffic.
Even with this I think by looking at the QMUL network tests that
most major sites don't have network capacity issues, but some
smaller sites may well be less lucky.
|