home *** CD-ROM | disk | FTP | other *** search
- 1 - 4. _N_e_w__F_e_a_t_u_r_e_s This chapter describes MPT features released in MPT 1.6.1 and later releases. Text that applies to IRIX systems only is marked "IRIX systems only." Text that applies to Linux systems only is marked "Linux systems only." 4.1 _N_e_w__F_e_a_t_u_r_e_s__f_o_r__M_P_T__R_e_l_e_a_s_e__1_._6_._1 This section describes the features that are new in the MPT 1.6.1 release. The MPT 1.6.1 release is a Linux only release, so all of the features described in this section apply to Linux systems. However, some of the features will be available also on IRIX systems in a subsequent release. These are marked, "IRIX systems, subsequent release." The features not marked are already available on IRIX systems and are being introduced for Linux systems with this release. 4.1.1 _G_e_n_e_r_a_l_i_z_e_d _R_e_q_u_e_s_t_s (_I_R_I_X _s_y_s_t_e_m_s, _s_u_b_s_e_q_u_e_n_t _r_e_l_e_a_s_e) The following MPI-2 generalized request routines are implemented for Linux in MPT 1.6.1: MMMMPPPPIIII____RRRReeeeqqqquuuueeeesssstttt____ggggeeeetttt____ssssttttaaaattttuuuussss MMMMPPPPIIII____GGGGrrrreeeeqqqquuuueeeesssstttt____ccccoooommmmpppplllleeeetttteeee MMMMPPPPIIII____GGGGrrrreeeeqqqquuuueeeesssstttt____ssssttttaaaarrrrtttt MMMMPPPPIIII____SSSSttttaaaattttuuuussss____sssseeeetttt____ccccaaaannnncccceeeelllllllleeeedddd MMMMPPPPIIII____SSSSttttaaaattttuuuussss____sssseeeetttt____eeeelllleeeemmmmeeeennnnttttssss Also, the MMMMPPPPIIII____SSSSTTTTAAAATTTTUUUUSSSS____IIIIGGGGNNNNOOOORRRREEEE or MMMMPPPPIIII____SSSSTTTTAAAATTTTUUUUSSSSEEEESSSS____IIIIGGGGNNNNOOOORRRREEEE symbol can be passed to any MPT routine in place of a scalar or an array, respectively, of type MMMMPPPPIIII____SSSSttttaaaattttuuuussss to indicate that no status information need be returned. 4.1.2 _R_O_M_I_O__1_._2_._4__(_I_R_I_X__s_y_s_t_e_m_s_,__s_u_b_s_e_q_u_e_n_t__r_e_l_e_a_s_e_) The 1.6.1 version of MPT includes ROMIO version 1.2.4 for Linux systems. New features in this release include aggregation control hints ccccbbbb____ccccoooonnnnffffiiiigggg____lllliiiisssstttt, rrrroooommmmiiiioooo____ccccbbbb____rrrreeeeaaaadddd and rrrroooommmmiiiioooo____ccccbbbb____wwwwrrrriiiitttteeee. XFS filesystems are supported under both IRIX and Linux operating systems. 4.1.3 _S_t_a_r_t_u_p__I_n_i_t_i_a_l_i_z_a_t_i_o_n MPT now initializes during the loading of the MPI library, rather than waiting until the first MMMMPPPPIIII____IIIInnnniiiitttt call. This allows MPI to control the environment for optimizing purposes when used with other libraries, such as OpenMP. To - 2 - initialize MPI first, include the ----llllmmmmppppiiii command last in the list of libraries on the compile line. 4.1.4 _A_r_r_a_y__S_e_r_v_i_c_e_s__S_u_p_p_o_r_t By default on Linux systems, mmmmppppiiiirrrruuuunnnn uses Array Services to launch MPI jobs and clean up stray MPI processes upon abnormal job termination. If the mmmmppppiiiirrrruuuunnnn command displays a message that Array Services is not running, check that it is installed correctly, as follows: 1. Determine if the Array Services package is installed by invoking the rrrrppppmmmm ----qqqq ssssggggiiii----aaaarrrrrrrraaaayyyyssssvvvvccccssss command. If it is installed, refer to the aaaarrrrrrrraaaayyyy____sssseeeerrrrvvvviiiicccceeeessss, aaaarrrrrrrraaaayyyydddd....ccccoooonnnnffff, and aaaarrrrrrrraaaayyyydddd....aaaauuuutttthhhh man pages for proper configuration. 2. Determine if the daemon is running by invoking the following command: ppppssss ----eeeeffff |||| ggggrrrreeeepppp aaaarrrrrrrraaaayyyydddd If the daemon is not running, start it by invoking the following command as root: ////eeeettttcccc////iiiinnnniiiitttt....dddd////aaaarrrrrrrraaaayyyy ssssttttaaaarrrrtttt If the daemon is running, but you have made configuration changes, as root, restart it by invoking the following command: ////eeeettttcccc////iiiinnnniiiitttt....dddd////aaaarrrrrrrraaaayyyy rrrreeeessssttttaaaarrrrtttt 3. Determine if your configuration contains any errors by invoking the following command: ////uuuussssrrrr////eeeettttcccc////aaaasssscccchhhheeeecccckkkk Remember, after making configuration changes, you will need to restart the aaaarrrrrrrraaaayyyydddd daemon, as previously noted. Note: On IRIX systems, mmmmppppiiiirrrruuuunnnn also uses Array Services by default. 4.1.5 _M_P_I_-_2__A_t_t_r_i_b_u_t_e_s__(_I_R_I_X__s_y_s_t_e_m_s_,__s_u_b_s_e_q_u_e_n_t__r_e_l_e_a_s_e_) MPT 1.6.1 supports MPI-2 attributes, as described in section 8.8 of the MPI-2 standard. - 3 - 4.1.6 _S_i_n_g_l_e_-_C_o_p_y__M_o_d_e__S_u_p_p_o_r_t_e_d MPT 1.6.1 supports single-copy mode for MPI communication on Linux systems. This feature is supported for both single- host MPI jobs and MPI jobs running across partitions. The MMMMPPPPIIII____BBBBUUUUFFFFFFFFEEEERRRR____MMMMAAAAXXXX environment variable must be set to an appropriate value to enable single-copy transfers. This is disabled by default. On Linux, the single-copy feature uses the xpmem module to map memory from one MPI process onto another during job startup. The mapped areas include the static region, private heap and stack region. Single copy is supported for contiguous data types from any of the mapped regions. Memory mapping is enabled by default on Linux. To disable it, set the MMMMPPPPIIII____MMMMEEEEMMMMMMMMAAAAPPPP____OOOOFFFFFFFF environment variable. The xpmem kernel module must be installed on your system for single-copy mode to be supported. The xpmem module is released with the OS. 4.1.7 _M_P_I_-_2__O_n_e_-_S_i_d_e_d__C_o_m_m_u_n_i_c_a_t_i_o_n__S_u_p_p_o_r_t MPT 1.6.1 supports a subset of one-sided communication on Linux systems. The following MPI-2 functions are supported: MMMMPPPPIIII____WWWWiiiinnnn____ccccrrrreeeeaaaatttteeee MMMMPPPPIIII____WWWWiiiinnnn____ffffeeeennnncccceeee MMMMPPPPIIII____PPPPuuuutttt MMMMPPPPIIII____GGGGeeeetttt These functions are supported for both single-host MPI jobs and MPI jobs running across partitions. On Linux, the one-sided functions use the xpmem module to map memory from one MPI process onto another during job startup. The mapped areas include the static region, private heap, and stack region. The one-sided functions are supported for windows residing in any of the mapped regions. Memory mapping is enabled by default on Linux. To disable it, set the MMMMPPPPIIII____MMMMEEEEMMMMMMMMAAAAPPPP____OOOOFFFFFFFF environment variable. The xpmem kernel module must be installed on your system to support the MPI one-sided functions. The xpmem module is released with the OS. - 4 - 4.1.8 _X_P_M_E_M__D_r_i_v_e_r__S_u_p_p_o_r_t_e_d__f_o_r__P_a_r_t_i_t_i_o_n_e_d__S_y_s_t_e_m_s MPT 1.6.1 supports the XPMEM (cross partition) device driver for Linux systems. This feature allows MPI processes running on one partition to communicate with MPI processes running on a different partition via the NUMAlink network. The MPI/XPMEM driver for Linux supports the standard send/receive model in both buffered and single-copy modes, in addition to the MPI-2 one-sided functions (MMMMPPPPIIII____WWWWiiiinnnn____ccccrrrreeeeaaaatttteeee,,,, MMMMPPPPIIII____WWWWiiiinnnn____ffffeeeennnncccceeee, MMMMPPPPIIII____PPPPuuuutttt, and MMMMPPPPIIII____GGGGeeeetttt). The use of the XPMEM driver allows MPI to extend the cache coherency domain of the system to include all the partitions being used, up to a 512 process limit. This allows MPI to maintain very low latency and high bandwidth across partitioned systems. The MPI/XPMEM driver requires the xpmem kernel module to be installed on your system. The xpmem module is released with the OS. When running an MPI job across partitions, MPI automatically selects the XPMEM interconnect, if available. For more information, see "Default Interconnect Selection" in the MMMMPPPPIIII man page. If MPI is successful in configuring the XPMEM driver on a partitioned system, the following message appears at startup when using the ----vvvv (verbose) option: MMMMPPPPIIII:::: UUUUssssiiiinnnngggg tttthhhheeee XXXXPPPPMMMMEEEEMMMM NNNNUUUUMMMMAAAAlllliiiinnnnkkkk iiiinnnntttteeeerrrrccccoooonnnnnnnneeeecccctttt.... 4.1.9 _M_P_I___A_l_l_o_c___m_e_m__a_n_d__M_P_I___F_r_e_e___m_e_m__S_u_p_p_o_r_t_e_d The MPT 1.6.1 release supports the MPI-2 functions MMMMPPPPIIII____AAAAlllllllloooocccc____mmmmeeeemmmm and MMMMPPPPIIII____FFFFrrrreeeeeeee____mmmmeeeemmmm on Linux systems. MMMMPPPPIIII____AAAAlllllllloooocccc____mmmmeeeemmmm allocates memory from the private heap, and is used in conjunction with MMMMPPPPIIII____FFFFrrrreeeeeeee____mmmmeeeemmmm, which frees the dynamically allocated memory. 4.1.10 _N_u_m_a___P_l_a_c_e_m_e_n_t___S_u_p_p_o_r_t Users have used environment variables to enable nonuniform memory access (NUMA) optimization placement on IRIX systems. Some of this functionality is now available on Linux systems. You can now use the MMMMPPPPIIII____DDDDSSSSMMMM____CCCCPPPPUUUULLLLIIIISSSSTTTT to link MPI ranks to the specified CPUs. You can also use the MMMMPPPPIIII____DDDDSSSSMMMM____MMMMUUUUSSSSTTTTRRRRUUUUNNNN environment variable when running on a dedicated system or within a restricted set of CPUs (either with ccccppppuuuusssseeeettttssss or rrrruuuunnnnoooonnnn) to ensure the MPI ranks are linked to those CPUs. - 5 - Other environment variables now supported on Linux systems are MMMMPPPPIIII____DDDDSSSSMMMM____OOOOFFFFFFFF and MMMMPPPPIIII____DDDDSSSSMMMM____VVVVEEEERRRRBBBBOOOOSSSSEEEE. 4.2 _N_e_w__F_e_a_t_u_r_e_s__f_o_r__M_P_T__R_e_l_e_a_s_e__1_._7 This section describes the features that are new in the MPT 1.7 release. For IRIX customers, the features in the previous section ("New Features for MPT Release 1.6.1) that are marked, "IRIX systems, subsequent release," are also new for this release. 4.2.1 _M_P_I___W_i_n___l_o_c_k__a_n_d__M_P_I___W_i_n___u_n_l_o_c_k__S_u_p_p_o_r_t_e_d The MPT 1.7 release supports the MPI-2 functions MMMMPPPPIIII____WWWWiiiinnnn____lllloooocccckkkk and MMMMPPPPIIII____WWWWiiiinnnn____uuuunnnnlllloooocccckkkk. These functions allow the user to protect access to remote memory windows. 4.2.2 _R_e_d_u_c_e_d__S_t_a_r_t_u_p__T_i_m_e__O_v_e_r_h_e_a_d Changes have been made to significantly reduce the startup time for MPI and SHMEM jobs. For more detailed information, see the iiiinnnnttttrrrroooo____sssshhhhmmmmeeeemmmm man page. 4.2.3 _S_H_M_E_M__S_u_p_p_o_r_t_e_d__o_n__L_i_n_u_x__S_y_s_t_e_m_s The MPT 1.7 release supports the SHMEM programming model for Linux systems. On Linux, mmmmppppiiiirrrruuuunnnn is required to launch SHMEM programs. For this release, SHMEM is supported for single host jobs only. For more detailed information, see the iiiinnnnttttrrrroooo____sssshhhhmmmmeeeemmmm man page. 4.2.4 _M_P_I___N_A_P__E_n_v_i_r_o_n_m_e_n_t__V_a_r_i_a_b_l_e The MPT 1.7 release supports the MMMMPPPPIIII____NNNNAAAAPPPP environment variable to control the way in which ranks wait for events to occur. You can find more information in the MMMMPPPPIIII(1) man page. 4.3 _N_e_w__F_e_a_t_u_r_e_s__f_o_r__M_P_T__R_e_l_e_a_s_e__1_._8 4.3.1 _S_u_p_p_o_r_t__f_o_r__H_I_P_P_I_-_8_0_0__O_S__B_y_p_a_s_s__W_i_l_l__C_e_a_s_e__i_n__M_P_T__1_._9 The OS bypass support for MPI across HIPPI will be removed from MPT feature releases after MPT 1.8. The MPT engineering team will focus future efforts on supporting MPI features on NUMAlink, GSN, and Myrinet-based systems and clusters. - 6 - 4.3.2 _S_u_p_p_o_r_t__f_o_r__M_P_I___I_N___P_L_A_C_E The MPT 1.8 release supports the MPI-2 argument MMMMPPPPIIII____IIIINNNN____PPPPLLLLAAAACCCCEEEE. Collective communications can occur "in place" for intracommunicators, with the output buffer being identical to the input buffer. This is specified by providing a special argument value, MMMMPPPPIIII____IIIINNNN____PPPPLLLLAAAACCCCEEEE, instead of the send buffer or the receive buffer argument. 4.3.3 _S_u_p_p_o_r_t__o_f__M_P_I___A_c_c_u_m_u_l_a_t_e The MPT 1.8 release supports the MPI-2 function MMMMPPPPIIII____AAAAccccccccuuuummmmuuuullllaaaatttteeee. This is a one-sided communication function that combines the data moved to the target process with the data that already resides at that process, rather than replacing the data there. 4.3.4 _A_d_d_i_t_i_o_n_a_l__N_U_M_A__p_l_a_c_e_m_e_n_t__s_u_p_p_o_r_t__o_n__L_i_n_u_x NUMA placement can be critical for MPI codes on Linux. This is also the case for hybrid OpenMP/MPI codes. Support has been added for "containing" the OpenMP threads near the MPI ranks in this release. Although these threads are not "pinned," they are restricted to run only on a certain set of CPUs, based on the OOOOMMMMPPPP____NNNNUUUUMMMM____TTTTHHHHRRRREEEEAAAADDDDSSSS environment variable. The "pinning" of the OpenMP threads will be considered in a future release. In addition, support for the MMMMPPPPIIII____DDDDSSSSMMMM____PPPPPPPPMMMM environment variable is enabled. The environment variable MMMMPPPPIIII____DDDDSSSSMMMM____MMMMUUUUSSSSTTTTRRRRUUUUNNNN has been deprecated in favor of MMMMPPPPIIII____DDDDSSSSMMMM____DDDDIIIISSSSTTTTRRRRIIIIBBBBUUUUTTTTEEEE on Linux only. The displayed information generated by the MMMMPPPPIIII____DDDDSSSSMMMM____VVVVEEEERRRRBBBBOOOOSSSSEEEE environment variable has been enhanced to display information related to these new features. 4.3.5 _E_n_h_a_n_c_e_d__M_P_I___B_a_r_r_i_e_r__a_n_d__M_P_I___W_i_n___f_e_n_c_e An optimized implementation of the MMMMPPPPIIII____BBBBaaaarrrrrrrriiiieeeerrrr and MMMMPPPPIIII____WWWWiiiinnnn____ffffeeeennnncccceeee functions has been introduced in this release. This optimized path can be used with barriers or fences on arbitrary communicators and windows. In this release, this optimized implementation is generally not activated by default. The feature can be activated by setting the MMMMPPPPIIII____BBBBAAAARRRR____DDDDIIIISSSSSSSSEEEEMMMM or MMMMPPPPIIII____BBBBAAAARRRR____TTTTRRRREEEEEEEE environment variables. For further details on activating this feature, see the MMMMPPPPIIII man page. 4.3.6 _I_m_p_r_o_v_e_d__M_P_I__L_a_t_e_n_c_y Changes have been made in the MPT 1.8 release to reduce the MPI send/receive latency. Several of these changes affect - 7 - both IRIX and Linux implementations. The most pronounced improvements can be seen on SGI Altix 3000 systems for send/receive latency measured between two processes that share the same SHub, and for latency between processes residing on different partitions. 4.3.7 _S_H_M_E_M__S_u_p_p_o_r_t__f_o_r__P_a_r_t_i_t_i_o_n_e_d__L_i_n_u_x__S_y_s_t_e_m_s The MPT 1.8 release supports the SHMEM programming model for partitioned Linux systems. On Linux, the mmmmppppiiiirrrruuuunnnn multiple host syntax is required to launch a SHMEM application on multiple partitions. SHMEM routines can be used exclusively in an application, or can be used in conjunction with MPI message passing routines in the same application. For more detailed information, see the iiiinnnnttttrrrroooo____sssshhhhmmmmeeeemmmm man page. 4.3.8 _S_t_a_c_k__T_r_a_c_e_b_a_c_k__a_n_d__C_o_r_e__D_u_m_p__C_o_n_t_r_o_l With MPT 1.8, if a rank aborts, either by calling MPI_Abort() or by receiving an unhandled signal which normally results in a core dump, MPT will display a stack traceback showing the location of the abort. MPT will limit creation of core files to the first rank on each host to abort, by default. More information is located in the MMMMPPPPIIII(1) man page under the MMMMPPPPIIII____CCCCOOOORRRREEEEDDDDUUUUMMMMPPPP and MMMMPPPPIIII____CCCCOOOORRRREEEEDDDDUUUUMMMMPPPP____DDDDEEEEBBBBUUUUGGGGGGGGEEEERRRR environment variable descriptions. 4.3.9 _R_O_M_I_O__M_o_d_i_f_i_c_a_t_i_o_n_s With MPT 1.8, the ROMIO software (which supplies the MPI-2 File I/O routines) has been modified so that the special MMMMPPPPIIIIOOOO____RRRReeeeqqqquuuueeeesssstttt datatype and the MMMMPPPPIIIIOOOO____WWWWaaaaiiiitttt() and MMMMPPPPIIIIOOOO____TTTTeeeesssstttt() routines are no longer required to be used. Instead, the MPI-1 MMMMPPPPIIII____RRRReeeeqqqquuuueeeesssstttt datatype and the MMMMPPPPIIII____WWWWaaaaiiiitttt****() and MMMMPPPPIIII____TTTTeeeesssstttt****() series of routines can be used. Existing binaries will continue to work properly if they call the older MMMMPPPPIIIIOOOO**** routines. Source code does not need to be modified to use the MMMMPPPPIIII____RRRReeeeqqqquuuueeeesssstttt datatype and routines. Your existing source will compile without comment. However, if you do modify your source and recompile, you will be able to call MMMMPPPPIIII____WWWWaaaaiiiittttaaaannnnyyyy(), for example, with a mix of MMMMPPPPIIII____RRRReeeeqqqquuuueeeesssstttt requests returned from communication calls like MMMMPPPPIIII____IIIIsssseeeennnndddd() and requests returned from I/O routines like MMMMPPPPIIII____FFFFiiiilllleeee____iiiiwwwwrrrriiiitttteeee(). This unification of I/O and communication requests can lead to simplification of your code and enhanced overlapping of I/O and program execution. - 8 - However, if you do modify your source as outlined above, you must execute the program using MPT 1.8 or later only; earlier versions of the MPT library will not correctly run such programs. 4.4 _N_e_w__F_e_a_t_u_r_e_s__f_o_r__M_P_T__R_e_l_e_a_s_e__1_._9 4.4.1 _M_P_I_-_2__P_r_o_c_e_s_s__M_a_n_a_g_e_r__I_n_t_e_r_f_a_c_e__F_u_n_c_t_i_o_n_a_l_i_t_y MPI-2 functionality for the MMMMPPPPIIII____CCCCoooommmmmmmm____ssssppppaaaawwwwnnnn and related functions and attributes described in sections 5.3, 5.5.1, and 5.5.3 of the MPI-2 standard is enabled for Linux partitioned systems in this release. For further details, see the mmmmppppiiiirrrruuuunnnn(1) and MMMMPPPPIIII(1) man pages. 4.4.2 _U_S_E _M_P_I Available for Linux Users of the Intel efc Fortran Compiler The USE MPI statement has been available in IRIX Fortran since MPT 1.4. With MPT 1.9, this feature has been added for users of efc on Linux as well. See MMMMPPPPIIII(1) man page for details, including the need for a -I option to efc. 4.4.3 _E_n_h_a_n_c_e_d _M_P_I__B_a_r_r_i_e_r _a_n_d _M_P_I__W_i_n__f_e_n_c_e _b_y _D_e_f_a_u_l_t _o_n _L_i_n_u_x In MPT 1.8, an enhanced barrier algorithm was implemented for the MMMMPPPPIIII____BBBBaaaarrrrrrrriiiieeeerrrr and MMMMPPPPIIII____WWWWiiiinnnn____ffffeeeennnncccceeee functions. This optimized algorithm uses the Altix hardware fetch-op synchronization registers, and scales as O(log(n)) as the number of processors increases. In MPT 1.9, this barrier algorithm is activated by default on Altix systems when the MMMMPPPPIIII____BBBBaaaarrrrrrrriiiieeeerrrr and MMMMPPPPIIII____WWWWiiiinnnn____ffffeeeennnncccceeee functions are called. 4.4.4 _A_d_d_i_t_i_o_n_a_l__N_U_M_A__P_l_a_c_e_m_e_n_t__S_u_p_p_o_r_t__o_n__L_i_n_u_x MMMMPPPPIIII____DDDDSSSSMMMM____DDDDIIIISSSSTTTTRRRRIIIIBBBBUUUUTTTTEEEE has now been enhanced to support multiple executables specified on the same mmmmppppiiiirrrruuuunnnn command. In addition support has been added so that MMMMPPPPIIII____DDDDSSSSMMMM____DDDDIIIISSSSTTTTRRRRIIIIBBBBUUUUTTTTEEEE also works for MMMMPPPPIIII____CCCCoooommmmmmmm____ssssppppaaaawwwwnnnn. This support will ensure that when run on a dedicated machine or within a cpuset that users can get optimal performance. Users of LSF 5.1 that use dynamic cpusets will cause MMMMPPPPIIII____DDDDSSSSMMMM____DDDDIIIISSSSTTTTRRRRIIIIBBBBUUUUTTTTEEEE to be set automatically. 4.4.5 _S_i_n_g_l_e_-_C_o_p_y__M_o_d_e__E_n_a_b_l_e_d__b_y__D_e_f_a_u_l_t MPT 1.9 now supports single-copy mode for MPI communication by default for certain situations. The eligible MPI functions include MMMMPPPPIIII____IIIIsssseeeennnndddd, MMMMPPPPIIII____SSSSeeeennnnddddrrrreeeeccccvvvv, MMMMPPPPIIII____AAAAllllllllttttooooaaaallllllll, MMMMPPPPIIII____BBBBccccaaaasssstttt, MMMMPPPPIIII____AAAAllllllllrrrreeeedddduuuucccceeee and MMMMPPPPIIII____RRRReeeedddduuuucccceeee. Users that use - 9 - these functions and transfer 2000 bytes or more of data will use the single-copy mode for the transfer. Users of MMMMPPPPIIII____SSSSeeeennnndddd should continue to use the MMMMPPPPIIII____BBBBUUUUFFFFFFFFEEEERRRR____MMMMAAAAXXXX environment variable to enable single copy. Users who wish to disable this optimization can set the MMMMPPPPIIII____DDDDEEEEFFFFAAAAUUUULLLLTTTT____SSSSIIIINNNNGGGGLLLLEEEE____CCCCOOOOPPPPYYYY____OOOOFFFFFFFF environment variable. 4.4.6 _M_P_I _C++ _b_i_n_d_i_n_g_s _f_o_r _I_n_t_e_l _C++ _8._0 _a_n_d _g++ _3._0 _c_o_m_p_i_l_e_r_s _o_n _L_i_n_u_x MPT 1.9 now contains two separate MPI C++ binding libraries. The default libmpi++.so (-lmpi++) library is compatible with code generated by the g++ 2.0 versions of the compiler but is not binary compatible with g++ 3.0 versions. A new library is now also available libg++3mpi++.so (-lg++3mpi++) that is binary compatible with code generated by the g++ 3.0 compilers as well as that produced by the Intel 8.0 C++ compiler. 4.4.7 _H_I_P_P_I_-_8_0_0__I_n_t_e_r_c_o_n_n_e_c_t__N_o__L_o_n_g_e_r__S_u_p_p_o_r_t_e_d The MPT 1.9 release no longer supports the HIPPI-800 OS bypass interconnect on IRIX systems. Support for any remaining HIPPI interconnect users will be limited to patches for MPT 1.8. 4.4.8 _M_P_I__S_u_p_p_o_r_t__f_o_r__L_a_r_g_e__A_l_t_i_x__S_y_s_t_e_m_s The MPT 1.9 release has been enhanced to allow support for more than 200 MPI processes on a single system image for Altix systems. Please see the FFFFiiiilllleeee____DDDDeeeessssccccrrrriiiippppttttoooorrrr____LLLLiiiimmmmiiiittttssss section in Chapter 6 of these Release Notes for information on how to increase the system limit of open file descriptors. 4.4.9 _g_7_7__P_r_o_f_i_l_i_n_g__E_n_t_r_y__P_o_i_n_t_s With the MPT 1.9 release, profiling entry points were added for Fortran codes using the g77 compiler on Altix. Profiling entry points are discussed in Section 4.18 of the MPI-2 Standard, and in Chapter 8 of the MPI-1.1 Standard. 
