1. Department of Computer Science Joshua Goehner and Dorian Arnold University of New Mexico (In collaboration with LLNL and U. of Wisconsin

2.  Key statistics from Top500 (Nov. 2011) ◦ 149/500 (30%) more than 8,192 core  In 2006, this number was 12/500 (2.4%) ◦ 7 systems: more than 64K cores ◦ 6 systems: between 64K and 128K cores ◦ 3 systems: more than 128K cores  Later this year: LLNL’s Sequoia w/ 1.6M cores  2018?: Exascale systems w/ 10 7 or 10 8 cores A Little Preaching to the Choir Software systems must scale as well!

3.  Before bootstrapping: ◦ Program image on storage device ◦ Set of (allocated) computer nodes  After bootstrapping:  Application processes started on computer nodes  Application’s configuration complete  ready for primary operation Software Infrastructure-bootstrapping Given a node allocation, start infrastructure's composite processes and propagate necessary startup information.

4. “I gotta get my application up and running!”

5. (1) “First, I start all the processes on the appropriate nodes”

6. (2) “Next, I must disseminate some initialization information”

7. Bootstrapping is complete when the infrastructure is ready for steady-state usage.

8. Basic Bootstrapping  Sequential instantiation ◦ e.g. rsh or ssh-based  Point-to-point propagation 70 seconds to start 2K processes!

9.  Infrastructure-specific, scalable mechanisms ◦ Still limited by sequential operations ◦ Not portable to other infrastructures  Using high-performance resource managers ◦ Myriad interfaces ◦ No communication facilities  Generic bootstrapping infrastructures ◦ LaunchMON: targets tools with wrapper for existing RMs ◦ ScELA: sequentially launch agents that create local processes More Scalable Approaches

10. MRNet Sequential-ish Bootstrapping  Parent creates children  Local  fork ()/ exec ()  Remote  rsh -based mechanism  Integrated instantiation and information propagation  MRNet’s “standard”

11. MRNet XT (hybrid) process launch  Bulk-launch 1 process per node  Process launches collocated processes  Information propagated in similar fashion

12. LIBI Approach  LIBI: Lightweight infrastructure- bootstrapping infrastructure ◦ Name is no longer a work in progress ◦ Generic service for scalable distributed software infrastructure bootstrapping  Process launch  Scalable, low-level collectives Large Scale Distributed Software Job Launchers Communication Services LIBI rsh/ssh SLURM OpenRTE ALPS LaunchMON DebuggersSystem Monitors Applications Performance Analyzers Overlay Networks COBO MPI

13.  session: set of processes (to be) deployed ◦ session master manages other members ◦ session front-end interacts with session master ◦ LIBI currently supports only master/member communication  host-distribution: where to create processes ◦  process distribution: how/where to create processes ◦ LIBI API

14. launch(process-distribution-array) ◦ instantiate processes according to input distributions [send|recv]UsrData(session-id, msg) ◦ communicate between front-end and session master broadcast(), scatter(), gather(), barrier() ◦ communicate between session master and members LIBI API (cont’d)

15. front-end( ){ LIBI_fe_init(); LIBI_fe_createSession(sess); proc_dist_req_t pd; pd.sessionHandle = sess; pd.proc_path = get_ExePath(); pd.proc_argv = get_ProgArgs(); pd.hd = get_HostDistribution(); LIBI_fe_launch(pd); //test broadcast and barrier LIBI_fe_sendUsrData(sess1, msg, len ); LIBI_fe_recvUsrData(sess1, msg, len); //test scatter and gather LIBI_fe_sendUsrData(sess1, msg, len ); LIBI_fe_recvUsrData(sess1, msg, len); return 0; } Example LIBI Front-end

16. session_member() { LIBI_init(); //test broadcast and barrier LIBI_recvUsrData(msg, msg_length); LIBI_broadcast(msg, msg_length); LIBI_barrier(); LIBI_sendUsrData(msg, msg_length) //test scatter and gather LIBI_recvUsrData(msg, msg_length); LIBI_scatter(msg, sizeof(rcvmsg), rcvmsg); LIBI_gather(sndmsg, sizeof(sndmsg), msg); LIBI_sendUsrData(msg, msg_length); LIBI_finalize(); } Example LIBI-launched Application

17.  LaunchMON-based runtime ◦ Tested SLURM or rsh launching ◦ COBO PMGR service LIBI Implementation Status Large Scale Distributed Software Job Launchers Communication Services LIBI rsh/ssh SLURM OpenRTE ALPS LaunchMON DebuggersSystem Monitors Applications Performance Analyzers Overlay Networks COBO MPI

18. LIBI Early Evaluation

19. LIBI Microbenchmark Results

21.  MRNet uses LIBI to launch all MRNet processes ◦ Parse topology file and setup/call LIBI_launch()  Session front-end gathers/scatters startup information ◦ Parent listening socket (IP/port) MRNet/LIBI Integration

22.  STAT over MRNet over LIBI performance results  Basic, scalable rsh-based mechanism  Hybrid (XT-like) launch approach  Mechanisms to alleviate filesystem contention ◦ Like our scalable binary relocation service  More flexible process and host distributions ◦ Instantiating different images in same session ◦ Integrating allocating and launching  Integrated scalable communication infrastructure Some Future Work