COSMOLOGICAL SIMULATIONS Modelling the Formation and Evolution of Cosmic Structure H.M.P. COUCHMAN Department of Physics and Astronomy University of Western Ontario London, ON N6A 3K7 Canada Abstract. Studies of the formation of cosmic structure are entering a period of dramatic improvements in sophistication and verisimilitude. A wealth of observational data is becoming available, ranging from studies of galaxies, clusters and their cosmic distribution, to the high redshift universe and observations of microwave background uctuations. These observations, together with enormous improvements in the scale and quality of the numerical computations that can be undertaken, promise very substantial progress in our understanding of the formation and evolution of cosmic structure. Numerical simulations are the crucial theoretical probe of non-linear gravitational and hydrodynamic cosmic evolution. The talk briey reviews the current state of numerical structure formation studies and discusses the developments in techniques which, when coupled with current supercomputers, are leading to studies of unprecedented scope and resolution. The talk concludes with comments on future directions for structure formation studies and on the challenges inherent in simulations of increasing size. Initial results from some of the rst billion particle simulations are presented. 1. Introduction Numerical simulations of cosmic structure have a number of goals. The ultimate aim of many investigations is to explore and understand the formation, distribution and morphology of galaxies. Convincing modelling of the astrophysical processes which contribute to the formation of galaxies is still 2 H.M.P. COUCHMAN some way in the future. Nevertheless, the simulation of very many aspects of cosmic structure has matured to a point where they are providing reliable tools for investigating processes occurring in the post-recombination universe. In particular, key aims are to distinguish world models, to constrain the initial uctuation spectrum and to understand non-linear gravitational evolution and hydrodynamics in the cosmic context. Traditionally, cosmological simulations have been concerned with the modelling of Large-Scale Structure, the distribution of voids, walls, clusters etc. and have been used to generate low order spatial and velocity statistics for a variety of cosmologies and initial uctuation spectra. New instruments and observations are now providing a wealth of data spanning many dierent epochs and spatial scales. The talk will outline the role that simulation plays in interpreting these observations in cosmology and review the progress that has been made to date as well as providing a few pointers to the future. 2. The Post-Recombination Universe The post recombination universe is shown schematically in Figure 1. Structure grows in the standard model from small uctuations present at the time of recombination through the so-called \Dark Ages" to the formation of the rst bound radiating objects. In standard hierarchical models structure grows through the aggregation of small objects into larger structures. Some of the key observational elements that must be described by a successful structure formation model are shown in Figure 1. Over the last decade our ability to probe to higher redshifts has increased dramatically with detailed observations of the Lyman forest becoming available as well as very deep surveys which promise the elusive goal of detecting primval galaxies. In the early 1980's, the initial spectrum was constrained only within very broad limits and the distribution of galaxies and clusters was known at low redshift. Numerical structure formation studies bridged this large redshift gap with collisionless simulations to map out the general features of gravitational clustering. The detailed measurement of the uctuation spectrum from studies of the microwave background early next decade, together with precise mapping of the galaxy distribution on large scales and the probing of structure to high redshifts will dramatically alter the nature of the gap that must be bridged by numerical simulation. With well constrained initial conditions, and a detailed understanding of the evolution of the collisionless component, the focus of many numerical studies will shift to careful investigation of the crucial hydrodynamic aspects of cosmic structure formation. 3 COSMOLOGICAL SIMULATIONS Redshift, z Cosmic Microwave Background 3 10 Recombination The "Dark Ages" . 30 . . . . . . . . . . . ❇ 3 ❇ ❇ 1 . . . . . . . . . . . . ❇ ❇ First Objects (Pop III stars?) ❇ Quasars Lyman alpha Galaxies X-ray clusters Large-Scale Structure, Hubble Deep Field etc. 0 Figure 1. Schematic comoving look-back cone. 3. The Growth of Structure Figure 2 illustrates the primary features of the growth of structure. The spectrum illustrated is typical of \CDM-like" initial conditions in which structure forms hierarchically, growing from small scales to large. For uctuations on some scale with amplitude less than unity the growth is linear; once the amplitude of a uctuation exceeds unity, the evolution is non-linear and numerical simulation is essential to accurately describe the structure. Numerical simulations thus have a fundamental role to play in connecting the initial linear spectrum with the non-linear regime. The further complication which hydrodynamic simulations are now beginning to address is the connection between the observed distribution of luminous matter with that of the gravitationally dominant dark matter distribution. On scales of galaxies and below, hydrodynamic simulation is an essential component of investigations of cosmic structure and of our interpretation of observational data. 4 H.M.P. COUCHMAN Bias: b(R)? NON-LINEAR LINEAR z=0 Galaxies (Observed) Growth and Evolution Dark Matter Growth ~x1000 z=1000 Galaxies Large Scale Structure Clusters COBE Figure 2. A schematic view of post-recombination growth of the uctuation spectrum, here represented as the mass variance in spheres of radius . Whilst the mass variance is less than unity, growth is linear; for larger amplitudes reliable prediction requires numerical simulation. R 4. Requirements of a Numerical Simulation Two features of gravitational collapse in the cosmic context are the presence of irregular geometries and large density contrasts. A bound object virialises with a density contrast of roughly 200 and this overdensity grows as the object maintains a constant physical size within the expanding background. A dissipated baryonic component will have a larger contrast. Lagrangian particle methods are ideally suited to accurately modelling these features. Indeed, it is dicult to see how one might easily model a collisionless component without representing it by particles. A number of techniques have been used to model the hydrodynamic component. Smoothed Particle Hydrodynamics (Monaghan 1992) is a fully Lagrangian method which ts well with gravitational particle techniques. Eulerian methods are also becoming popular in the cosmic context, but, to be successful, adaptive mesh rene- COSMOLOGICAL SIMULATIONS 5 ment (AMR) techniques must be employed in order to faithfully resolve hydrodynamic phenomena over a wide range of densities. The range of scales present in observed cosmic structures is very large; from sub-galactic scales to superclustering and beyond, far greater than can be modelled in a single simulation. The particle number that can be modelled sets the mass resolution of the simulation. A simple example will outline the scale of the problem. Suppose that we were to tackle the \Grand Challenge" of modelling the distribution of galaxies in a representative sample of the universe. A cubic region of the universe 200 Mpc on a side would contain of order 105 galaxies. Requiring a minimum of 100 mass elements (or particles) per galaxy and supposing that 10% of the total matter resides in galaxies (including their haloes) would suggest a simulation with 108 particles. Simulations of this size are only now becoming feasible. Useful hydrodynamic modelling of galaxies in the same context would require a resolution perhaps 10 times larger. This sort of resolution is considerably greater than that which can be achieved in hydrodynamic simulations at present. Whilst by no means all simulation programmes span this range of scales, the argument is suggestive of the mass resolution that is desirable and explains the continual push towards higher particle number by practitioners in the eld. 5. Achievements of cosmological simulations The argument of the previous section suggests that very high particle number is desirable in cosmological simulations. Figure 3 illustrates the remarkable progress that has been achieved in improving the number of particles that can be simulated. The gure shows a representative sample of leading-edge or near leading-edge simulations in the particle-number stakes at various epochs, as well as indicating the algorithmic method used to compute inter-particle forces and the general type of computer hardware used. Particle number is only one aspect of the problem of course; dierent algorithms have dierent memory requirements and compute the various types of problem with diering eciency. Furthermore the hardware groupings are very coarse; no mention has been made of vector computers for example. Nonetheless, a dramatic improvement in attainable resolution over time is apparent as is the necessity of using parallel supercomputers to achieve the highest resolution. Numerical simulations of structure formation are modelling simple physical laws. The key to progress is understanding the collective phenomena which are modelled. Table 1 lists several areas in cosmology which have been addressed by numerical simulation and assesses the degree of success or reliability with which the various problems have been modelled. 6 H.M.P. COUCHMAN Figure 3. Sketch showing the particle number achieved in various state-of-the-art simulations for dierent algorithms as a function of epoch. The representative simulations illustrated were run either on serial computers, parallel computers (typically massively parallel systems) or using the special purpose GRAPE hardware. TABLE 1. Achievement and reliability of cosmological simulations. Achievement Gravitational: Checks of perturbation theory Exploring validity of simple analytic models (e.g., Press{Schechter) Non-linear gravitational evolution, stable clustering etc. Distinguishing dierent initial , spectra, , cluster abundances etc. Comparison of predicted gg with observed distributions; bias etc. Hydrodynamic: Cluster gas and properties Lyman Galaxy formation, rst objects etc. Reliability Good|symbiotic relationship Reasonable|models limited Reasonable|need greater dynamic range Good Rudimentary|need reliable galaxy identication techniques|hydro Good Fairly good|low order statistics used so far Rudimentary|need cooling and to model star formation, feedback etc. COSMOLOGICAL SIMULATIONS 7 6. The Current State-of the-Art The continuing push for higher resolution has lead to an increasing use of massively parallel supercomputers for leading edge simulations. These computers oer both the processing speed|provided the algorithms can be eciently parallelized|as well as providing the large total memory needed for storage of the particle data and for the computational overhead of the algorithm itself. Despite the diculty of programming for distributed memory architectures, parallel versions of several N-body algorithms exist; in particular, versions of Tree, P3 M and AP3 M algorithms have been implemented. Impressive performance is possible on current hardware. The Virgo (1996) collaboration uses an AP3 M algorithm which has been parallelised using Cray's proprietary data-parallel language CRAFT, to run on the T3D supercomputer. We can achieve a performance of roughly 2,500 particle updates/second/processor with excellent scaling behaviour up to 256 processors for a 17 Million particle run (for details see Pearce & Couchman 1997). The same code but without adaptive renements (essentially P3 M) has been implemented using Cray's \shmem" library|an MPI2like one-sided message passing paradigm|for the Cray T3E (MacFarland et al. 1998). The performance achieved in this case is 8,000 particle updates/second/processor. A 2563 (17 Million) particle run of approximately 2000 steps takes 10 hours on 128 processors of a Cray T3E. This is impressive performance even considering the slow down that occurs with the non adaptive P3 M implementation as clustering develops. Storage requirements become equally impressive as the particle number increases. For collisionless simulations 6 words are required per particle or roughly 10 words for hydrodynamic simulations. Each output slice of a 2563 particle collisionless simulation, for example, requires 500 Mb for storage of positions and velocities in 32 bit precision. Typically 10 to 15 output times are stored per simulation. Two broad approaches are being pursued to achieve higher resolution. Selective volume renormalization allows very high resolution to be achieved in the investigation of galaxies, clusters or patches of the universe at high redshift for example. With proper care it is possible to model the largescale cosmological eld by using nested shells of variable mass particles and/or the application of external elds. Where uniform resolution is required over the whole computational domain there is little alternative to the straightforward calculation with very large particle number. Such a simulation has been undertaken recently by the Virgo collaboration. A cube of side 4000 Mpc containing 109 particles was evolved from a start redshift of 29. A slice of the nal output of the simulation at redshift zero together 8 H.M.P. COUCHMAN with the physical parameters used is shown in Figure 4. The code used was the shmem P3 M code described above running on 512 processors of the T3E in Garching, Germany. The full simulation of 500 steps took 70 hours. Each output is 25Gb with a total dataset of 1=3 Tb. With this size dataset, analysis is a challenge, with much of the post-processing, of necessity, being done on the the T3E or large machines local to it. 7. Conclusions There are two broad conclusions to this talk. First, the promise of secure knowledge of the initial uctuation spectrum together with improved reliability of collisionless simulations suggests that in the next few years| barring upsets to the overall cosmological picture and current understanding of structure formation|we will have a good understanding of the evolution and structure of the dark matter distribution. This will provide the basis from which we can undertake detailed hydrodynamical simulations of the formation of galaxies and cosmic structure. Observations at high redshift are providing detailed information which will place signicant constraints on the details of structure formation models as realised by numerical simulations. The second conclusion is computational. As is evident from the data presented above, the attainable resolution has increased by 4 orders of magnitude in under two decades. With the advent, and increasing use, of massively parallel computers the simulation of very large numbers of particles is becoming feasible. This allows us to model a range of scales where studies|particularly hydrodynamic studies|of very many aspects of cosmic structure will begin to have detailed predictive power. References MacFarland, Tom, Couchman, H.M.P., Pearce, F.R. and Pichlmeier, Jakob, (1998) A New Parallel P3 M Code for Very Large-Scale Cosmological Simulations, New Astronomy, submitted (astro-ph/9805096) Monaghan, J.J. (1992) Annual Reviews of Astronomy and Astrophysics, 30, 543 Pearce, F.R. and Couchman, H.M.P. (1997) Hydra: An parallel adaptive grid code, New Astronomy, v 2 n 5, p 411 Virgo: Colberg, J.M., Couchman, H.M.P., Efstathiou, G.P., Frenk, C.S., Jenkins, A., Nelson A.H., Peacock, J.A., Pearce, F.R., Thomas, P. A. and White, S.D.M. (1996) 9 COSMOLOGICAL SIMULATIONS = 1; , = 0:21; h = 0:5; 8 = 0:6 CDM Main slice: 20002 20 h,3 Mpc3 Enlargement: 450 240 20 h,3 Mpc3 P3 M: z = 29; s = 100 h,1 kpc; 10003 particles; 10243 mesh; Cray T3E|512 cpus Mparticle = 2 1012 h,1 M i 200 Mpc/h 50 Mpc/h Figure 4. A slice through the 109 particle simulation recently completed by the Virgo consortium. The cosmology simulated is standard Cold Dark Matter but with a modied spectrum (shape parameter, , = 0 21). :