Scalable Compression and Replay of Communication Traces

Characterizing the communication behavior of larges cale ap p lications is a d iffi cult and cos tly tas k d ue to cod e/s y s tem comp lex ity and their long ex ecution times . A n alternative to running actual cod es is to gather their communication traces and then rep lay them, w hich facilitates ap p lication tuning and future p rocurements . W hile p as t ap p roaches lack ed los s les s s calable trace collection, w e contribute an ap p roach that p rovid es ord ers of magnitud e s maller, if not near cons tant-s ize, communication traces regard les s of the number of nod es w hile p res erving s tructural information. W e introd uce intraand inter-nod e comp res s ion techniq ues of M P I events and p res ent res ults of our imp lementation for B lueG ene/L . G iven this novel cap ability , w e d is cus s its imp act on communication tuning and bey ond . T o the bes t of our k now led ge, s uch a concis e rep res entation of M P I traces in a s calable manner combined w ith d eterminis tic M P I call rep lay are w ithout any p reced ence.