The paper discusses the cosmological simulation code GADGET-2, a massively parallel TreeSPH code capable of simulating collisionless fluids using the N-body method and ideal gases with smoothed particle hydrodynamics (SPH). The code conserves energy and entropy in regions free of dissipation and allows for adaptive smoothing lengths. Gravitational forces are computed using a hierarchical multipole expansion, with an optional TreePM algorithm that combines short-range forces computed via the tree method and long-range forces determined by Fourier techniques. Time integration is based on a quasi-symplectic scheme, allowing for different timesteps for long-range and short-range forces. The domain decomposition for parallelization is based on a space-filling curve, ensuring high flexibility and tree force errors independent of domain cutting methods. The code is efficient in terms of memory consumption and communication bandwidth, having been used to perform large-scale cosmological N-body and SPH simulations, including over $10^{10}$ dark matter particles and over 250 million particles with radiative cooling and star formation. The paper presents the algorithms used, their accuracy, and performance through various test problems, and highlights the code's public release to the research community.The paper discusses the cosmological simulation code GADGET-2, a massively parallel TreeSPH code capable of simulating collisionless fluids using the N-body method and ideal gases with smoothed particle hydrodynamics (SPH). The code conserves energy and entropy in regions free of dissipation and allows for adaptive smoothing lengths. Gravitational forces are computed using a hierarchical multipole expansion, with an optional TreePM algorithm that combines short-range forces computed via the tree method and long-range forces determined by Fourier techniques. Time integration is based on a quasi-symplectic scheme, allowing for different timesteps for long-range and short-range forces. The domain decomposition for parallelization is based on a space-filling curve, ensuring high flexibility and tree force errors independent of domain cutting methods. The code is efficient in terms of memory consumption and communication bandwidth, having been used to perform large-scale cosmological N-body and SPH simulations, including over $10^{10}$ dark matter particles and over 250 million particles with radiative cooling and star formation. The paper presents the algorithms used, their accuracy, and performance through various test problems, and highlights the code's public release to the research community.