Hi all
Maybe I didn’t look in the correct place, but I couldn’t find any details in the documentation about the initial setup of the OS for EOS hosts.
We want to deploy a shared EOS instance for the VOs using CTA. Obviously for the RAL Tier-1 we will want it to appear on the LHCOPN subnet (and LHCONE), while other users would expect something different.
Do EOS hosts have any issue handling multiple public (IPv4) addresses and would the CTA backend host have any issue talking to EOS in this situation?
Thanks
Alastair
Hi Alastair,
I think we should first understand why you want a single EOS instance - certainly that’s not how things are run at CERN and there are benefits to multiple instances, especially if you want to differentiate the networking. Remember that multiple EOS instances can communicate with the same central CTA instance behind.
Apart from that, if you don’t insist that the disk servers (FSTs) are shared then things will be easier, i.e. a single FST is either on the OPN or not. EOS was happily managing disk servers on different networks (e.g. between Wigner and Meyrin).
Oliver.
The main reason to have a single EOS instance is the benefit of a shared buffer, which can cope with a big request from one user as the other users are generally not busy. We support all the LHC VOs as well as Dune and a number of other smaller Grid VO in addition to numerous local users. We will certainly need some shared instances.
Many users have small and/or spikey usage but when averaged together its a fairly predicatble amount.
In Castor we had a 1.5PB buffer shared by everyone, provided by 10 - 15 storage nodes. With EOS we have 12 storage nodes each with 16 x 2TB drives for storage (separate SSDs for OS and an NVMe for quarkDB). We can create separate instances, but the hardware is not infinitely divisbale and for VOs like ALICE where we are pledged to provide only 3% of their resources, its hard to justify dedicated hardware.
We could just give the EOS nodes an IP address in the OPN subnet and it would probably be fine for now, but we want to leave as many options open as possible.
Alastair