Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.stecf.org/conferences/adass/adassVII/beccianiu.html
Дата изменения: Mon Jun 12 18:51:51 2006
Дата индексирования: Tue Oct 2 00:24:05 2012
Кодировка:
Поисковые слова: п п р п р п р п р п р п р р п п р п п р п п р п п р п п р п

Table of Contents -- Index -- PS reprint -- PDF reprint

Astronomical Data Analysis Software and Systems VII
ASP Conference Series, Vol. 145, 1998
Editors: R. Albrecht, R. N. Hook and H. A. Bushouse

Parallel Tree N-body Code: Data Distribution and DLB on the CRAY T3D for Large Simulations

U.  Becciani,   V.  Antonuccio-Delogu,   M.  Gambera  and  A.  Pagliaro
Osservatorio Astrofisico di Catania, Città Universitaria, Viale A. Doria, 6 - I-95125 Catania - Italy

R.  Ansaloni
Silicon Graphics S.p.A. St.6 Pal. N.3, Milanofiori I-20089 Rozzano (MI) - Italy

G.  Erbacci

Cineca, Via Magnanelli, 6/3 I-40033 Casalecchio di Reno (BO) - Italy

 

Abstract:

We describe a strategy for optimal memory and work distribution. We have performed a series of tests to find an optimal data distribution in the Cray T3D memory, and to identify a strategy for the Dynamic Load Balance (DLB). The results of tests show that the step duration depends on two main factors: the data locality and the network contention. In a very large simulation, due to network contention, an unbalanced load arises. To remedy this we have devised an automatic work redistribution mechanism which provided a good DLB.

           

1. Introduction

N-body simulations are one of the most important tools in contemporary theoretical cosmology, however the number of particles required to reach a significant mass resolution is more larger than those allowed even by present-day state-of-the-art massively parallel (Gouhing 1995; Romeel 1997 & Salmon 1997) supercomputers (hereafter MPP). The most popular algorithms are generally based on grid methods like the P³M. The main problem with this method lies in the fact that the grid has typically a fixed mesh size, while the cosmological problem is inherently highly irregular. On the other hand the Barnes & Hut (1986, hereafter BH) oct-tree recursive method is inherently adaptive, and allows one to achieve a higher mass resolution. Because of these features, the computational problem can easily become unbalanced and cause performance degradation. For this reason, we have undertaken a study of the optimal work- and data-sharing distribution for our parallel treecode.

Our Work- and Data-Sharing Parallel Tree-code (hereafter WDSH-PTc) is based on this algorithm tree scheme, which we have modified to run on a shared-memory MPPs (Becciani, 1996). The BH-Tree algorithm is a NlogN procedure to compute the gravitational force, a more detailed discussion on the BH tree method can be found in Barnes & Hut (1986). For our purposes, we can distinguish three main phases in each timestep: TREE_FORMATION (TF), FORCE_COMPUTE (FC), UPDATE_POSITION. Besides, in the FC phase we can distinguish two important subphase: TREE_INSPECTION (TI) and ACCEL_COMPONENTS (AC).

2. The WDSH-PTc

During the FC phase each PE computes the acceleration components for each body in asynchronous mode and only at the end of the phase an explicit barrier statement is set. Our results show that the most time-consuming phases (TF, TI and AC) are executed in a parallel regime. Tests were carried out, fixing the constraint that each PE executes the FC phase only for all bodies residing in the local memory. A bodies data distribution ranging from contiguous blocks (coarse grain: CDIR$ SHARED POS(:BLOCK,:)) to a fine grain distribution (tf) (CDIR$ SHARED POS(:BLOCK(1),:)) was adopted. We studied different tree data distributions ranging from assigning to contiguous blocks (tc) a number of cells equal to the expected number of internal cells (N_Tcell), [Salmon 1990](coarse grain: CDIR$ SHARED POS_CELL(BLOCK(:N_TCELL/N$PES),:)), to a simple fine grain distribution (CDIR$ SHARED POS_CELL:BLOCK(1),:)), (tf).

All the tests were performed for two different set of initial conditions, namely uniform and clustered distribution having 2²⁰ particles each and they were carried out using from 16 to 128 PEs. In Tab. 1 we report only the most significant results obtained with 128 PEs.

 

PE# p/sec FC phase T-step UF

1Mun_tf_bf 128 4129 230.05 249.5 4.22

1Mcl_tf_bf 128 3832 250.32 268.81 4.57

1Mun_tf_bm 128 3547 270.51 290.45 5.90

1Mcl_tf_bm 128 3308 291.63 312.26 6.32

1Mun_tf_bc 128 4875 186.31 205.32 4.14

1Mcl_tf_bc 128 4490 203.37 222.72 4.38

1Mun_tc_bc 128 837 1051.93 1230.0 16.33

1Mcl_tc_bc 128 750 1173.24 1373.4 17.62

Our results show that the best tree data distribution is obtained using a block factor equal to 1, then we can to conclude that a (tf) should be used for the kind of codes that we deal in this paper.

The fine grain bodies data distribution (bf) is obtained using a block factor N=1; i.e., bodies are shared among the PE but there is no spatial relation in the body set residing in the same PE local memory. The medium grain bodies data distribution (bm) is obtained using a block factor $N=N_{bod}/ 2*N\$PEs$; i.e., each PE has two data block of bodies properties residing in the local memory, each block having a close bodies set. At the end the coarse grain bodies data distribution (bc) is obtained using a block factor $N=N_{bod}/N\$PEs$; i.e., each PE has one close data set block of bodies residing in the local memory. The results reported in Tab. 1 show that the best bodies data distribution, having the highest code performance in terms of particles per second, is obtained using the block factor $N=N_{bod}/N\$PEs$ as expected, due to the data locality effect.

3. Dynamic Load Balance

Here, we present the results of a new DLB strategy, that allows us to avoid any large overhead. The total time spent in a parallel region T_tot, can be considered as the sum of the following terms

T_tot=T_s+ K T_p/p + T_o(p)

where p is the number of processors executing the job, T_s is the time spent in the serial portion of the code, T_p is the time spent by a single processor (p=1) to execute the parallel region, T_o(p) the overhead time due to the remote data access and to the synchronization points, and K is a constant.

 

Figure 1: a) 64 PE run; b) 1 Million of particles: homogeneous configuration

In the FC phase, there are no serial regions, so $T_{p} \propto$ to the length of the interaction list (IL). Using a coarse grain subdivision, each PE has a block of close bodies in the local memory ($Np=N_{bod}/N\$PEs$); in a uniform distribution initial condition, the PEs having extreme numeration in the pool of available PEs, have a lower load at each timestep. This kind of effect may be enhanced, if a clustered initial condition is used. If the number of PEs involved in the simulation increases we note that the data dispersion on the T3D torus increases. Our results do not show (see Figure 1a) the existence of a relationship between the time spent in the FC phase and the total length of the IL. The adopted technique is to perform a load redistribution among the PEs so that all PEs have the same load in the FC phase. We force each PE to execute this phase only for a fixed portion of the bodies residing in the local memory NB_lp. NB_lp is given by

$$NB_{lp}=(N_{bod}/N\$PEs)*P_{lp}$$ (2)

being P_lp= const. ($0 \leq P_{lp} \leq 1$). The FC phase for all the remaining bodies

$$N_f=N\$PEs*(N_{bod}/N\$PEs)*(1 - P_{lp})$$ (3)

is executed from all the PEs that have finished the FC phase for the NB_lp bodies. No correlation between the PE memory and the PE, executing the FC phase for it, is found. If P_lp=1 all PEs execute the FC phase only for bodies residing in the local memory, on the contrary if P_lp =0, N_f= N_bod (NB_lp = 0), the PEs execute only N_f bodies and the locality is not taken into account. Several tests were performed with different values of P_lp. We report in Figure 1b, only, the case with 10⁶ particles uniform.

The results obtained lead us that it is possible to fix a P_lp value allowing the best code performances. In particular, we note that is convenient to fix the P_lp value near to , that is maximize the load balance. The data show that, fixing the PEs number and the particles number, the same P_lp value gives the best performance both in uniform and clustered conditions. This means that it is not necessary to recompute the P_lp value to have good performances.

4. Final Considerations

The results obtained using the WDSH-PTc code, at present, give performances comparable to those obtained with different approaches such as Local Essential Tree (LET) (Dubinski 1996), with the advantage of avoiding the LET and an excessive demand for memory. Besides, a strategy for the automatic DLB has been described, which does not introduce a significant overhead.

The results of this work will allow us to obtain, in the next future, a WDSH-PTc version for the CRAY T3E system, using the HPF-CRAFT and the shmem library. The new version will include an enhanced grouping strategy and periodic boundary conditions (Gambera & Becciani 1997).

References:

Gouhing, X., 1995, ApJ. Supp., 97, 884

Romeel, D.,& Dubinski, J., Hernquist, L., 1997, New Astronomy, 2, 277

Salmon, J., 1997, Proc. of the 8th SIAM Conf.

Barnes, J., & Hut, P., 1986, Nature, 324, 446

Becciani, U., Antonuccio-Delogu, V., & Pagliaro, A., 1996, Comp. Phys. Com., 99, 9

Salmon, J., 1990, Ph.D. Thesis, Caltech

Dubinski, J., 1996, New Astronomy, 1, 133

Gambera, M., & Becciani, U., 1997, in preparation

© Copyright 1998 Astronomical Society of the Pacific, 390 Ashton Avenue, San Francisco, California 94112, USA