GASNet gemini-conduit and aries-conduit documentation Larry Stewart User Information: ----------------- This documentation covers both gemini-conduit (for the Cray XE and XK series) and aries-conduit (for the Cray XC). The terms "Gemini" and "gemini-conduit" are used throughout this document when both gemini and aries are meant. Because these Cray systems have non-trivial differences between the login nodes and the compute nodes, one must use a cross-configure script to build either of these conduits. See other/contrib/cross-configure-crayxe-linux for Gemini and other/contrib/cross-configure-crayxc-linux for Aries. Where this conduit runs: ----------------------- The gemini-conduit runs on systems with Cray's "Gemini" interconnect, including their XE and XK series machines. There is no known minimum version required of any hardware or software component, but it is likely that sufficiently old releases of the system software are not supported correctly. The aries-conduit runs on systems with Cray's "Aries" interconnect, including their XC series machines. To the best of our knowledge, there are no known compatibility issues with any release of the XC series system software. Configuring job launch: ---------------------- Both ALPS (srun) and native SLURM (srun) are supported for job-launch by the provided gasnetrun_gemini and gasnetrun_aries scripts. The value of the environment variable MPIRUN_CMD can be used to choose between these two. The value when configure is run establishes a default, which a value set at job-launch time can override. Our recommended settings are + For ALPS (default): MPIRUN_CMD="aprun -n %N %C" + For native SLURM: MPIRUN_CMD="srun -K0 -W60 %V -mblock -n %N %C" For SLURM-based systems, one can simply set APRUN="srun -K0 -W60 %V -mblock" in ones environment when running the appropriate cross-configure script. If you desire additional customization, please see mpi-conduit's README for documentation regarding the MPIRUN_CMD environment variable. Note, that it is NOT required that one compile mpi-conduit to use gemini or aries conduits. Recognized environment variables: --------------------------------- * All the standard GASNet environment variables (see top-level README) GASNET_BARRIER - barrier algorithm selection In addition to the algorithms in the top-level README, there is an implementation specific to Gemini and Aries: GNIDISSEM - like RDMADISSEM, but implemented using lower-level GNI operations for lower latency. Currently GNIDISSEM is the default on both Gemini and Aries. GASNET_NETWORKDEPTH_SPACE - volume of AM Request queue This determines the maximum volume of AM Requests (in bytes) that can be oustanding between a given pair of peers before flow-control. This differs from most other conduits in which the depth is given by a count. Values smaller than 2 maximum size Medium AMs, or larger than 64 maximum sized Medium AMs, will be silenly adjusted to that range. The default value is 12K. GASNET_NETWORKDEPTH_TOTAL - depth of AM Reply queue This determines the maximum number of AM Requests that can be oustanding from one peer to all others before flow-control. The default value is 64, and the minimum is 1. GASNET_GNI_NUM_PD - number of post descriptors AM and RDMA operations all requires use of a control block called a Post Descriptor. This variable controls how many of these are allocated, and therefore limit the total number of operations which can be in-flight at any time. The default value is 512, and the minimum is 32. GASNET_GNI_GET_FMA_RDMA_CUTOVER - default 4096 max 16384 GASNET_GNI_PUT_FMA_RDMA_CUTOVER - default 4096 max 16384 These set the cutover from Fast Memory Access (RDMA commanded in user mode) to RDMA (commanded through the kernel). Transfers up to and including these sizes will be performed using FMA. Use of FMA will typically result in lower latency, but at the expense of greater involvement by the CPU (and thus less potential, for instance, for overlap of non-blocking communication with computation). GASNET_GNI_GET_BOUNCE_REGISTER_CUTOVER - default 4096, max 32768 GASNET_GNI_PUT_BOUNCE_REGISTER_CUTOVER - default 4096, max 32768 Every Post Descriptor has a small bounce buffer (128 bytes) for handling out-of-segment transfers (normally FMAs) with the least overhead. Beyond that size out-of-segment can be performed using either bounce buffers or dynamic registration. Such transfers up to and including these sizes will be conducted using bounce buffers. Beyond these sizes, dynamic memory registration is used. The minimum allowed is GASNET_PAGESIZE (normally 4K) and lower values will be silently replaced by this value. GASNET_GNI_BOUNCE_SIZE - default 256K, space for bounce buffers There is a pool of bounce buffers, each of a fixed size determined by the larger of GASNET_GNI_{PUT,GET}_BOUNCE_REGISTER_CUTOVER (see above). The value of GASNET_GNI_BOUNCE_SIZE determines the size of the entire pool of bounce buffers. Thus the number of bounce buffers available is GASNET_GNI_BOUNCE_SIZE / MAX(GASNET_GNI_PUT_BOUNCE_REGISTER_CUTOVER, GASNET_GNI_GET_BOUNCE_REGISTER_CUTOVER, (gasnet_AMMaxMedium() + 64)) Since all three terms in the MAX default to 4K, the default 256K gives 64 bounce buffers. If an operation must allocate a bounce buffer and the pool is empty, it will poll the completion queue until one is returned. Therefore, this setting may limit the number of operations requiring use of a bounce buffer which can be in-flight at any time. GASNET_GNI_MEM_CONSISTENCY This control the use of the flag GNI_MEM_STRICT_PI_ORDERING, GNI_MEM_RELAXED_PI_ORDERING, or neither when registering memory for RDMA. The legal values are the strings "strict", "relaxed" or "default". In this release the default is "relaxed". GASNET_GNI_MEMREG This controls the maximum number of concurrent temporary memory registrations, used for RDMA transfers to and from addresses outside of the GASNet segment. The default value is 16 on Gemini, and is 3072 on Aries. If this vale is too large (possibly due to competition with MPI or another runtime library) you will see a fatal error at startup. If GASNet detects the resource exhaustion, the message includes the following text: UDREG_CacheCreate() failed with rc=2 Alternatively, if the resource exhaustion is detected by the MPI runtime you may see a message including the following text: UDREG_CacheCreate unknown error return 2 Other runtimes may include similar failure messages. If you encounter these failures you should reduce the value of this parameter, or in the extreme you may wish to set GASNET_USE_UDREG=0 to disable use of UDREG by GASNet. GASNET_PHYSMEM_PINNABLE_RATIO - default 0.80 GASNET_PHYSMEM_MAX - default * GASNET_PHYSMEM_NOPROBE - default 0 The logic for setting the maximum size of the segment is copied from portals-conduit, which uses the following setting to control the maximum size of the segment. If you set PINNABLE_RATIO 1 (or higher) and PHYSMEM_MAX, and PHYSMEM_NOPROBE you will bypass all the checking. Of course mmap might fail, or you might be swapping the segment... GASNET_USE_UDREG AVAILABLE UNLESS CONFIGURED WITH --disable-gni-udreg. This gives a boolean: "0" to disable or "1" to enable the use of the UDREG library for caching of local memory registrations, and is intended primarily for debugging purposes. The default is enabled. GASNET_USE_FIREHOSE AVAILABLE ONLY IF CONFIGURED WITH --enable-gni-firehose. This gives a boolean: "0" to disable or "1" to enable the use of the firehose dynamic pinning library for local memory registration, and is intended primarily for debugging purposes. The default is enabled. GASNET_DOMAIN_COUNT GASNET_GNI_PTHREADS_PER_DOMAIN GASNET_AM_DOMAIN_POLL_MASK AVAILABLE ONLY IF CONFIGURED WITH --enable-gni-multi-domain. For their use, please see the multi-domain documentation under the "Optional configure-time settings" section, below. PMI_GNI_LOC_ADDR - NIC address on local node PMI_GNI_COOKIE - Access code shared by all nodes in job PMI_GNI_PTAG - Protection tag shared by all nodes in job PMI_GNI_DEV_ID - Device ID (Selects which NIC is in use) Provided by Cray runtime system. Don't mess with these. Optional configure-time settings: ------------------------------ --with-gni-max-medium=[value] By default gasnet_AMMaxMedium() is 4032: a 4096 byte buffer minus 64 bytes reserved for up to 16 handler arguments. This configure option allows control over the value of gasnet_AMMaxMedium(). The value must be a multiple of 64, and cannot be less than 512. The default value is 4032. --enable-gni-multi-domain This configure option enables EXPERMIENTAL support for improved performance of RDMA operations in a PAR (multi-threaded) build of Gemini or Aries conduit by allocating multiple GNI Communication Domains and distributing the clients threads over these domains. Since operations on a given domain must be serialized (we use a pthread mutex), use of multiple domains can significantly reduce lock contention. The trade-offs are that with multiple domains the maximum size of the GASNet segment may be reduced, and progress on Active Messages may be slowed. The following three environment variables are honored when multi-domain support is enabled: GASNET_DOMAIN_COUNT This is the number of Communication Domains to create per process. The multi-domain support has no benefits unless this is set to a value larger than 1. The default value is 1 (a single domain with no benefits). GASNET_GNI_PTHREADS_PER_DOMAIN This is the number of threads to assign to the first domain before assigning any to the second. When the last domain has been assigned this many threads, then assignment resumes at the first domain. The default value is 1 (cyclic/round-robin assignment). GASNET_AM_DOMAIN_POLL_MASK This value controls how often threads in domains other than the first poll for arrival of Active Messages (threads in the first domain poll on every call to gasnet_AMPoll). Since all Active Message traffic arrives via the first domain, polling by more than one thread will increase lock contention, but a total lack of polling by threads in other domains can lead to deadlock in some circumstances. A value of 0 will result in polling for Active Messages on every call (explicit or implict) to gasnet_AMPoll(). Values of 1, 3, 7, 15, etc. will result in progressively lower average frequencies (1 in (n+1)). The default is 255 (1 in every 256 polls by threads outside the first domain will poll for AM arrivals). --enable-gni-firehose This configure option enables EXPERMIENTAL support for the firehose dynamic memory registration library for caching of dynamic memory registration. This is only caching of local registration, and NOT the "full" firehose which provides cached remote registration (as for GASNET_SEGMENT_EVERYTHING support). NOTE: This option is largely superseded by the (ON by default) use of Cray UDREG, which does not suffer from firehose's vulnerability when memory is unmapped. Optional compile-time settings: ------------------------------ * All the compile-time settings from extended-ref (see the extended-ref README) in gemini-conduit/gasnet_gemini.h: GASNETC_USE_SPINLOCK - 0 user gasneti_mutex_t (default) 1 use gasneti_spinlock_t The _DEFAULT values for all of the GASNET_GNI_ environment variables. in gemini-conduit/gasnet_gemini.c: int gasnetc_poll_burst = 10; /* number of rdma completions to get at once */ Sets the number of events for arriving active messages taken at once #define AM_BURST 20 Known problems: --------------- * See the Berkeley UPC Bugzilla server for details on known bugs. Future work: ------------ * Caching of memory registrations (Firehose?) gemini-conduit uses a single fixed memory registration for the GASNet segment. It uses another for the AM receive buffers, and transiently uses additional registrations for large RDMA transfers outside the segment. There are only a modest number of mapping registers available per rank, so large numbers of registrations are inappropriate, but small numbers could be cached rather than per-rdma transient. * Inline It should be possible to inline the special cases of short blocking GET and short blocking PUT * Alternative designs The O(N) AM buffers queues could be used only for wakeup, leaving AM transport to RDMA, more along the lines of the shmem-conduit. This would reduce buffer requirements for large jobs. ============================================================================== Design Overview: ---------------- The gemini conduit provides the gasnet core and extended APIs on Cray systems based on the gemini interconnect. Initial work done on the XE6. References: Cray documentation "Using the GNI and DMAPP APIs" document S-2446-3102 http://docs.cray.com/cgi-bin/craydoc.cgi?mode=SideMap;f=xe_sitemap APIs: PMI-2 /opt/cray/pmi/default/include GNI /opt/cray/gni-headers/default/include/gni_pub.h ==Background== The Gemini interconnect was introduced with the Cray XE6 system. It may be thought of as an evolutionary development of the Seastar interconnect used on the Red Storm system and the Cray XT series. Physically Gemini is a 3D torus system using network interfaces that connect directly to the Hyperchannel IO bus of AMD processors. Logically Gemini provides short messages, remote memory access, remote atomic operations, and remote block transfers. Memory must be registered with GNI before it can be used for communications. Job launch services are provided by the Cray Application Level Placement Scheduler (ALPS), with API provided through implementations of the PMI (process management interface) API. During initialization, a parallel application creates a Communications Domain, which is basically a binding to a particular network interface. Data necessary to get permit ranks of a parallel application to communicate are provided via environment variables and PMI interfaces. In gemini, the address information for a remote rank is stored in a software structure called an Endpoint. Generally a separate endpoint is needed for each rank of the job. Each endpoint is associated with a Completion Queue (CQ) which signals when locally originated operations to a particular endpoint have finished. Cqs may be used by multiple endpoints. NOTE: The next paragraph is no longer accurate - we no longer use SMSG To use the short message system (SMSG), the application allocates buffer space, and assigns a completion queue to be used to announce the arrival of messages from remote ranks. Short message system messages can be of variable length, but each message uses a fixed length message buffer from a pool. The shared message queue (MSGQ) system is similar, except that resources are shared at the node level, rather than each rank having separate resources. The shared message queue system has better scalability, scaling as O( (N/24)**2 ) rather than O( N**2 ). However, MSGQ has a maximum message length of 128, while the limit for SMSG is large enough to handle a complete GASNet Medium Active Message. In ether SMSG or MSGQ, the buffer memory must be registered using the GNI API. The Gemini interconnect provides two flavors of RDMA. FMA, or Fast Memory Access, provides remote atomic operations, and remote get and put. These operations are commanded by user-mode applications directly to the hardware, (so called OS Bypass). For larger messages, Cray provides an RDMA facility based on a hardware Byte Transfer Engine, but its use is mediated by the kernel mode device driver. It isn't clear when one should use RDMA rather than FMA, but a few hundred bytes to a few thousand bytes is likely. ==Overview== The gemini conduit creates an endpoint for every rank, and then binds each endpoint to a particular rank using the address information exchanged through PMI. Each endpoint in gemini is also associated with a common completion queue, which is a structure used for announcing the completion of locally originated messages. Next the gemini conduit allocates and registers storage for buffers for the gemini short message system. The setup allocates buffers large enough to hold a complete AMMedium message. Arriving short messages are signaled on a second completion queue. Using defaults, about 32K per rank are reserved on each rank. ==Active Message design== The gemini conduit supports PSHM for same-node communications. This isn't strictly necessary when using the Cray Short Message System, but it would be necessary when using the Cray Shared Message Queue system because MSGQ cannot send to the same node. Flow control and deadlock avoidance is done by a fixed credit system. The environment variable GASNET_NETWORKDEPTH is the number of AM buffers for each peer. The default value is 12. Each peer has an am_credit variable. When a rank wishes to send an active message, it atomic decrements the peer's credit counter and calls the gemini conduit's Poll routine until it succeeds. At the receiver, every AM Request generates a "hidden" credit return message if and only if the handler did not send a Reply. NOTE: The Gemini short message system has some sort of internal credit accounting, but we don't rely on it because of lack of features (allocating space for replies) and lack of visibility (the only way to find out if there is credit is to try to send a message). The gemini-conduit credit design is such that the underlying GNI_SmsgSend should never fail. Short and Medium active messages take one message each, so there is no need for a reassembly queue for arriving fragments. Long active messages are more complex. If the data of a long AM will fit in a maximum sized message, then it is sent as immediate data and copied at the destination. If the data is too large for the AM packet, then it is sent by RDMA, and when the RDMA completion event occurs, the Active message is sent. For blocking AMs, the originator polls for completion of the RDMA by calling gasnetc_poll_local_queue, which checks for RDMA completions but does NOT check for arriving active messages. In the case of a non-blocking long active message, the active message packet is constructed in memory attached to the RDMA post descriptor, and when the RDMA completion event occurs, the AM is sent by the completion routine. ==RDMA== Within the Gemini API (GNI), RDMA is commanded by creating a Post Descriptor and passing it to GNI. Both the source and destination memory of an RDMA transfer must be registered. The knowledge of RDMA is encapsulated in the functions named gasnetc_rdma_{put,get}*(). GASNet requires that the remote address of a put or get will be within the GASNet memory segment, which is pre-registered by gemini-conduit. GASNet does not require that the local address of a put or get be registered. There are a number of cases. As mentioned above, Gemini provides two kinds of RDMA: FMA and BTE (called "RDMA"). FMA is Fast Memory Access and is used for short transfers and is commanded entirely in user mode. The Byte Transfer Engine is intended for larger transfers and is kernel mediated. Gemini-conduit allocates a pool of RDMA Post Descriptors, called gasnetc_post_descriptor. These are kept on a LIFO queue in registered memory, because each one contains an "immediate" bounce buffer to be used in the event the local address of a short xfer is not registered. Gemini-conduit also allocates a pool of medium sized bounce buffers, which are used in the event the local address of a transfer is not registered and is too large for the immediate buffer. For transfers to or from memory outside the GASNet segment that are too large for the medium sized bounce buffer, the local buffer is registered for the duration of the transfer, then unregistered. The post descriptor also contains a completion field, that defines what to do when the RDMA is finished, and a set of flags that control completion behavior. RDMA Completion event processing is done by gasnetc_poll_local_queue. A number of types of follow-on activities are implemented: * For a GET, potentially copy data from a bounce buffer to the actual target * Potentially De-register specially registered memory * Potentially free a dedicated bounce buffer * Potentially send an Active Message that was waiting for the RDMA completion * Perform completion processing (signal other software waiting for RDMA completion The GASNet Core API only uses RDMA PUT. The extended API uses both GET and PUT. The logic for GET is similar to that for PUT. ==Extended API== Gemini conduit uses the reference implementation of gasnet_op_t for completions. Due to occasional failures of memory registration of large blocks, if an RDMA transfer is larger than one megabyte, it is broken into smaller segments and all but the last are sent by blocking protocol. NOTE: The breakup of large blocks should probably be done for Long AMs as well, to allow increasing gasnet_AMMaxLong{Request,Reply}(). ==Locking== In order to support GASNET_PAR mode, a variety of locks are used. GASNETC_LOCK_GNI, GASNETC_UNLOCK_GNI The Cray gni API is not thread-safe, so the gemini conduit uses a single big lock around calls to gni. Certain cases might be thread safe, but without knowledge of the internals of gni it is hard to tell. Queue Locks A singly linked lists are used for the smsg_work_queue, and has its own lock. The are enqueue and dequeue functions that work on this queue. The bounce buffer pool and post descriptor pool use the gasneti_lifo_push and gasneti_lifo_pop functions. These are lock-free on the x86-64 architecture. ==Event Processing== The gemini conduit uses two gni completion queues: bound_cq_handle for completions of locally originated RDMA, and smsg_cq_handle, for arriving short messages. ===bound_cq_handle (RDMA completion)=== On a call to gasnetc_poll_local_queue, up to gasnetc_poll_burst (default 10) completions are processed. If the completion queue is empty, the function exits early. For each completed post descriptor, the flags are checked and post-completion activities performed: GC_POST_COPY - copy from bounce buffer to target GC_POST_SEND - send an AM that was waiting for PUT completion (AMLong) GC_POST_UNREGISTER - deregister one-time memory registration GC_POST_UNBOUNCE - free an allocated bounce buffer GC_POST_COMPLETION_FLAG - atomic set a variable to signal a waiting thread GC_POST_COMPLETION_OP - gasnete_op_markdone ===smsg_cq_handle=== NOTE: This section no longer accurate - we no longer use SMSG When a short message arrives, an event is sent to the completion queue associated with the endpoint of the message source. The receiver looks at the event to figure out which endpoint signaled, then calls GNI_SmsgGetNext(endpoint) to get the message. However, events are generated whenever a message arrives, but they may be out of order, and SmsgGetNext only delivers messages in-order. Consequently, when you get an event, there may appear to be no message waiting, or there may be several messages waiting. It is necessary to keep calling SmsgGetNext until you get all the waiting messages. At the same time, it is necessary to read the completion queue at least once for every message received or the completion queue will fill up. If the completion queue overflows, some events may be lost and it will be necessary to drain all endpoints to make sure no messages are left waiting. The work starts in gasnetc_poll_smsg_queue. First a call is make to gasnetc_poll_smsg_completion_queue, which reads up to AM_BURST events from the completion queue and quickly adds the relevant endpoints to a temporary array. Then, after releasing the GNI big lock, the array is scanned, and any peers which are not already on the smsg_work_queue are added to it. The smsg_work_queue is a list of endpoints which may have arriving traffic. Finally, gasnetc_poll_smsg_completion_queue checks to see if the smsg completion queue overflowed, in which case the completion queue is drained and ALL endpoints are added to the smsg_work_queue Then for up to SMSG_PEER_BURST ranks, gasnetc_process_smsg_q(source rank) is called to deal with any arriving messages. In gasnetc_process_smsg_q, arriving short messages are read and a dispatch is made on the message type. In the case of active messages, the message is copied into a local variable, the SMSG buffer is released, and then the AM handler is called. This is done in order to prevent stack growth caused by recursive calls to process arriving messages. Each time around the message processing loop, gasnetc_poll_smsg_completion_queue is called, which may result in additional ranks being queued on the smsg_work_queue, but does not cause recursive calls to gasnetc_process_smsg_q. This is done so that the average rate of retiring completion events is equal to the rate of processing messages, and prevents completion queue overflow. ==Exit protocol== NOTE: This section no longer accurate - we do more than is described here. The exit algorithm is borrowed from the portals-conduit. The Cray runtime system ALPS (Application Layout and Placement System) is strict about exiting without calling PMI_Finalize. It will interpret that as an error exit and kill all the other ranks. ==Initialization== The logic to determine the maximum size pinnable segment is borrowed from the portals-conduit. The gasnetc_bootstrap* functions containsthe interface to PMI. It also includes the implementation of AllGather for GASNet bootstrapExchange and GNI initialization. It uses PMI_AllGather. For some reason, PMI_AllGather does not return data in rank order, so it is sorted into the usual AllGather order internally. gasnet_core.c routine gasnetc_init also contains code to determine which ranks have shared memory. This is done using the PMI_Get_pes_on_smp() call, translated into the right format for gasnet_nodemap[]. The gasnet auxseg machinery is used to allocate registered memory for bounce buffers and for the rdma post descriptor pool. This is done to reduce the number of separately registered segments, because registrations are a scarce resource. == Build notes ==