| 19 | 19 | * 15:15 CZW: Mark raised an interesting question, and I wanted to get my understanding of the logic down someplace. We've added the new stsci1X.Y nodes into nebulous, and enabled them to accept data. This allows nebulous to use them as a target for things that do not have an explicit target set (replicated copies of data, files with no host request, files that have a host request that cannot be serviced due to the disk being full, etc). Nebulous is fully controlling these hosts. However, the majority of hosts have targetting rules set in pantasks, as part of the ippTasks/ipphosts.mhpcc.config file. This file sets rules that target OTAXY onto host ippXYZ.0, and do a similar thing with skycell_ids. Once these rules are defined, pantasks will begin to write data directly to those hosts. As this can cause NFS issues if pantasks attempts to write a large amount of data to a single host, we use the ~ipp/ippconfig/pantasks_hosts.input file to restrict this. Pantasks will not launch more than N jobs that rely on a given host, provided it knows about that host. To prevent this issue with the stsci0X.Y nodes, we add one instance of each host, and then turn off that instance. This prevents processing from happening on those hosts, but allows pantasks to manage the jobs that read/write from them. This means that if we add targeting to the stsci1X.Y nodes, we need to ensure that we add the hosts to pantasks_hosts.input to prevent overloading them. |
