Optimal processor count

The results presented in Section 6 suggest that all future work on NEMO should be carried out using code compiled with the PGI compiler suite as it gives the lowest runtimes.

The NOCS researchers ideally want to be able to run an entire model year, i.e. 365 model days, in a 12 hour run on HECToR as this enables them to make optimal use of the machine/queues and also allows them to keep up with the post-processing and data transfer of the results as the run progresses. They can currently achieve 300 model days in a 12 hour run using 221 processors. In this section we investigate whether an optimal processor count which satisfies the desire to complete a model year in a 12 hour time slot can be found. To do this NEMO is executed over a range of processors and the number of model days which can be computed in 12 hours, $ndays$, is obtained from:-

$$ndays = 43200 / t_{60}$$ (1)

where $43200$ is the number of seconds in 12 hours and $t_{60}$ is the time taken to complete a 60 step (i.e. 1 day) run of NEMO. This means we ideally need $t_{60} <= \frac{43200}{365} = 118.36$ seconds. The processor count investigated varies from 159 to 430. In all tests runs have been performed with the land cells removed. The results of this test are summarised in table 7. Figure 6 shows the results in graphical form with the 365 day threshold marked by the dashed line.

Table 7: Runtime for 60 time steps for various processor configurations ranging from 159 to 430.

jpni	jpnj	No. of procs	Time for 60 steps (seconds)
13	14	159	177.583
14	14	174	163.633
14	15	187	172.191
15	15	196	157.858
15	16	209	153.450
16	16	221	145.078
16	17	232	137.507
17	17	244	127.705
17	18	260	135.688
18	18	274	127.103
18	19	286	122.639
19	19	304	125.880
19	20	321	118.081
20	20	335	117.830
20	21	349	107.464
21	21	364	113.491
21	22	379(380)	114.175
22	22	398(396)	107.051
22	23	413	123.939
23	23	430(429)	110.871

Figure 6: Investigation of optimal processor count for NEMO subject to completing a model year within a 12 hour compute run. The dashed line shows the cut-off point.

In performing this investigation some problems were discovered relating to the computation of land only cells performed by the nocspmap_r25 code. It was found that several processor configurations yielded incorrect numbers of land cells. These have been highlighted in table 7 where the value which was incorrectly computed is given in ``()'' after the correct number of land cells. If the wrong number of land cells are specified the code fails with an error of the form:-

 ===>>> : E R R O R
         ===========

  Eliminate land processors algorithm
 
   jpni =           21  jpnj =           22                                                                       
                                                                                       
   jpnij =          380 < jpni x jpnj                                                                             
                                                                                       
 
   ***********, mpp_init2 finds jpnij=          379