We investigate the possibility of achieving high-fidelity universal two-qubit gates by supplementing optimal tuning of individual qubits with dynamical decoupling (DD) of local noise. We consider simultaneous local pulse sequences applied during the gate operation and compare the efficiencies of periodic, Carr-Purcell, and Uhrig DD with hard pulses along two directions ( pulses). We present analytical perturbative results (Magnus expansion) in the quasistatic noise approximation combined with numerical simulations for realistic noise spectra. The gate efficiency is studied as a function of the gate duration, of the number of pulses, and of the high-frequency roll-off. We find that the gate error is nonmonotonic in , decreasing as in the asymptotic limit, , depending on the DD sequence. In this limit -Urhig is the most efficient scheme for quasistatic noise, but it is highly sensitive to the soft UV cutoff. For small number of pulses, control yields anti-Zeno behavior, whereas pulses minimize the error for a finite . For the current noise figures in superconducting qubits, two-qubit gate errors , meeting the requirements for fault-tolerant quantum computation, can be achieved. The Carr-Purcell-Meiboom-Gill sequence is the most efficient procedure, stable for noise with UV cutoff up to gigahertz.