Scheduling and Load Balancing in the Parallel ROOT Facility (PROOF)

monitoring interface; concrete implementations for MySQL [8] and MonALISA [9] are already available. The PROOF basic scheduler implementation provides the possibility to assign the worker nodes based on the load and the group priority according to the simple formula: Nworkers = N f ree CPU · f · P̄j + Nmin where N f ree CPU is the number of free CPUs on the cluster nodes, f and Nmin are, respectively, the fraction of free units and the minimal number of units to assign to a job, and P̄j is the normalized priority previously defined. 3.3 Scheduling tests in ALICE: first results The mechanism for worker level resource control described above has been tested in the ALICE setup described in Sec. 2.2. For this purpose the job resource information was posted to the ALICE MonALISA repository and a dedicated daemon, provided by ALICE, was used to calculate the effective group priorities based on the required policy. The resulting priorities were then feed in the basic central scheduler and communicated to the active sessions. Figure 6: Results of CPU quota control test run by the ALICE experiment over one day. For each of the four groups the target (right) and achieved (left) CPU quotas are shown. The results of a CPU quota control test run by the ALICE experiment over one day is shown in fig. 6. For this test, four groups were defined and assigned quotas ranging from 10 to 40%. Users in the groups were continously running typical ALICE analysis jobs. As it can be seen in the plot, the target quotas are reached within few percentage units.