上一篇文章中说道,我们在 rte.h 中发现了有价值的说明:
我们一块一块来分析,首先看到第一块,关于 Process name Object:
* (a) Process name objects and operations // 进程名Object * 1. Definitions for integral types ompi_jobid_t and ompi_vpid_t. * The jobid must be unique for a given MPI_COMM_WORLD capable of * connecting to another OMPI_COMM_WORLD and the vpid will be the * process's rank in MPI_COMM_WORLD. * 2. ompi_process_name_t - a struct that must contain at least two integer-typed fields: * a. ompi_jobid_t jobid // 作业ID,应该也就是 MPI_COMM_WORLD,为了让不同的WORLD可以相连,每个WORLD就用这个作业ID来标识 * b. ompi_vpid_t vpid // 在 MPI_COMM_WORLD 的 进程rank * Note that the structure can contain any number of fields beyond these * two, so the process name struct for any particular RTE can be whatever * is desired. * 3. OMPI_NAME_PRINT - a macro that prints a process name when given // 打印进程名字的宏 * a pointer to ompi_process_name_t. The output format is to be * a single string representing the name. This function should * be thread-safe for multiple threads to call simultaneously. * 4. OMPI_PROC_MY_NAME - a pointer to a global variable containing * the ompi_process_name_t for this process. Typically, this is * stored as a field in the ompi_process_info_t struct, but that * is not a requirement. * 5. OMPI_NAME_WIlDCARD - a wildcard name. * 6. ompi_rte_compare_name_fields - a function used to compare fields * in the ompi_process_name_t struct. The function prototype must be * of the form: * int ompi_rte_compare_name_fields(ompi_rte_cmp_bitmask_t mask, * ompi_process_name_t *name1, * ompi_process_name_t *name2); * The bitmask must be defined to indicate the fields to be used * in the comparison. Fields not included in the mask must be ignored. * Supported bitmask values must include: * b. OMPI_RTE_CMP_JOBID * c. OMPI_RTE_CMP_VPID * d. OMPI_RTE_CMP_ALL * 7. uint64_t ompi_rte_hash_name(name) - return a string hash uniquely * representing the ompi_process_name passed in. * 8. OMPI_NAME - an Opal DSS constant for a handler already registered // 序列化,反序列化?DSS? * to serialize/deserialize an ompi_process_name_t structure.
第二块,是关于集体信息交换的:
* (b) Collective objects and operations // 集体对象 * 1. ompi_rte_collective_t - an OPAL object used during RTE collective operations // nodex 是指 要求每个进程“发现”作业中所有其他进程的相关互连联系信息。 * such as modex and barrier. It must be an opal_list_item_t and contain the * following fields: * a. id (ORTE type: int32_t) * b. bool active * flag that user can poll on to know when collective * has completed - set to false just prior to * calling user callback function, if provided * 2. ompi_rte_modex - a function that performs an exchange of endpoint information // 各节点进程间消息传输 * to wireup the MPI transports. The function prototype must be of the form: * int ompi_rte_modex(ompi_rte_collective_t *coll); * At the completion of the modex operation, the coll->active flag must be set * to false, and the endpoint information must be stored in the modex database. // nodex 的信息会存储到数据库 * This function must have barrier semantics across the MPI_COMM_WORLD of the * calling process. * 3. ompi_rte_barrier - a function that performs a barrier operation within the * RTE. The function prototype must be of the form: * int ompi_rte_barrier(ompi_rte_collective_t *coll); * At the completion of the barrier operation, the coll->active flag must be set * to false
更多的Modex操作信息,唯一能找到的参考是: https://github.com/open-mpi/ompi/wiki/ModexlessLaunch
启动MPI作业通常不仅要求各个进程在各个节点上生成,还要求每个进程“发现”作业中所有其他进程的相关互连联系信息。 完成后面这一步的默认启动机制被称为“modex”,包含几个步骤:
1.在启动时,每个进程打开每个接口驱动程序以查询本地节点的可用接口
2.那些具有接口的驱动程序会注册一个包含其接口联系信息的modex条目
3.作业中的流程执行集体操作以交换其个人联系信息。 这是在MPI_Init期间发生的阻塞操作。
也就是 MPI_Init 会进行一个各个进程间的信息交换,并且是有同步保障的。
第3块: 进程结构体
每个进程都有这样一个结构体,记录自己的 node 的 rank , 记录自己在 node 中的 rank 。
* (c) Process info struct * 1. ompi_process_info_t - a struct containing info about the current process. //当前进程的信息 * The struct must contain at least the following fields: * a. app_num - * b. pid - this process's pid. Should be same as getpid(). * c. num_procs - Number of processes in this job (ie, MCW) // 作业的进程数,一个作业可能包括多个节点,每个节点可能有多个进程 * d. my_node_rank - relative rank on local node to other peers this run-time // 节点 rank * instance knows about. If doing dynamics, this may be something * different than my_local_rank, but will be my_local_rank in a * static job. * d. my_local_rank - relative rank on local node with other peers in this job (ie, MCW) // 本节点上的 process rank * e. num_local_peers - Number of local peers (peers in MCW on your node) // 本地节点进程个数 * f. my_hnp_uri - * g. peer_modex - a collective id for the modex operation // 不知道 modex啥意思 * h. peer_init_barrier - a collective id for the barrier during MPI_Init * i. peer_fini_barrier - a collective id for the barrier during MPI_Finalize * j. job_session_dir - * k. proc_session_dir - * l. nodename - a string representation for the name of the node this * process is on * m. cpuset - * 2. ompi_process_info - a global instance of the ompi_process_t structure. * 3. ompi_rte_proc_is_bound - global boolean that will be true if the runtime bound * the process to a particular core or set of cores and is false otherwise.
第4块,初始化和反初始化操作:
* (e) Init and finalize objects and operations * 1. ompi_rte_init - a function to initialize the RTE. The function * prototype must be of the form: * int ompi_rte_init(int *argc, char ***argv); * 2. ompi_rte_finalize - a function to finalize the RTE. The function * prototype must be of the form: * int ompi_rte_finalize(void); * 3. void ompi_rte_wait_for_debugger(void) - Called during MPI_Init, this * function is used to wait for debuggers to do their pre-MPI attach. * If there is no attached debugger, this function will not block.
第5块,数据库操作:
* (f) Database operations * 1. ompi_rte_db_store - a function to store modex and other data in // Modex 的记录插入数据库 * a local database. The function is primarily used for storing modex * data, but can be used for general purposes. The prototype must be * of the form: * int ompi_rte_db_store(const ompi_process_name_t *proc, * const char *key, const void *data, * opal_data_type_t type); * The implementation of this function must store a COPY of the data * provided - the data is NOT guaranteed to be valid after return * from the call. * 3. ompi_rte_db_fetch - * NOTE: Fetch accepts an 'ompi_proc_t'. * int ompi_rte_db_fetch(const struct ompi_proc_t *proc, * const char *key, * void **data, * opal_data_type_t type); * 4. ompi_rte_db_fetch_pointer - * NOTE: Fetch accepts an 'ompi_proc_t'. * int ompi_rte_db_fetch_pointer(const struct ompi_proc_t *proc, * const char *key, * void **data, * opal_data_type_t type); * 5. Pre-defined db keys (with associated values after rte_init) * a. OMPI_DB_HOSTNAME * b. OMPI_DB_LOCALITY
其实这篇文章并没有给我们提供什么有实质性的信息,只是大致指明了一些方向。
还记起,我们在分析 MPI_Init 的消息时,并没有进入到 实际的 ompi_mpi_init 函数中,下一次,我们就要尝试进入该函数。
最简单的MPI程序就2行代码: MPI_Init(); MPI_Finalize;
初始化过程中应该是做了非常多的事情的,包括很多进程信息初始化,交换信息等,我们慢慢地去探索。