NAME

The GNI Fabric Provider

OVERVIEW

The GNI provider runs on Cray XC (TM) systems utilizing the user-space Generic Network Interface (uGNI) which provides low-level access to the Aries interconnect. The Aries interconnect is designed for low-latency one-sided messaging and also includes direct hardware support for common atomic operations and optimized collectives.

REQUIREMENTS

The GNI provider runs on Cray XC systems running CLE 5.2 UP04 or higher using gcc version 4.9 or higher.

When using the fi_mr_regattr() and fi_mr_regv() functions to register multiple region, users must register the memory region with 4K page alignment. Any other page address alignment will result in a return value of -FI_EINVAL.

SUPPORTED FEATURES

The GNI provider supports the following features defined for the libfabric API:

Endpoint types
The provider supports the FI_EP_RDM, FI_EP_DGRAM, FI_EP_MSG endpoint types, including scalable endpoints.
Address vectors
The provider implements both the FI_AV_MAP and FI_AV_TABLE address vector types. FI_EVENT is unsupported.
Memory registration modes
The provider implements the FI_MR_BASIC memory registration mode.
Data transfer operations
The following data transfer interfaces are supported for all endpoint types: FI_ATOMIC, FI_MSG, FI_RMA, FI_TAGGED. See DATA TRANSFER OPERATIONS below for more details.
Completion events
The GNI provider supports FI_CQ_FORMAT_CONTEXT, FI_CQ_FORMAT_MSG, FI_CQ_FORMAT_DATA and FI_CQ_FORMAT_TAGGED with wait objects of type FI_WAIT_NONE, FI_WAIT_UNSPEC, FI_WAIT_SET.
Modes
The GNI provider does not require any operation modes.
Progress
For both control and data progress, the GNI provider supports both FI_PROGRESS_AUTO and FI_PROGRESS_MANUAL, with a default set to FI_PROGRESS_AUTO. Note that for data progress, progression is only performed when data transfers use the rendezvous protocol.
Wait Objects
The GNI provider specifically supports wait object types FI_WAIT_UNSPEC, and FI_WAIT_SET. A wait object must be used when calling fi_cntr_wait, fi_cq_sread/from, fi_eq_sread/from, fi_wait. The GNI provider spawns an internal wait progress thread that is woken up when clients utilize the wait system (e.g., calling fi_wait).
Additional Features
The GNI provider also supports the following capabilities and features:
  • FI_MULTI_RECV
  • FI_SOURCE
  • FI_FENCE
  • FI_RM_ENABLED
  • FI_RMA_EVENT
  • FI_REMOTE_CQ_DATA
  • FABRIC_DIRECT compilation mode
  • FI_MORE (For FI_RMA)

DATA TRANSFER OPERATIONS

FI_ATOMIC

Currently, the GNI provider only supports atomic operations supported directly by the Aries NIC. These include operations on 32- and 64-bit, signed and unsigned integer and floating point values. Specifically,

Basic (fi_atomic, etc.)

  • FI_MIN, FI_MAX (no unsigned)
  • FI_SUM (no 64-bit floating point)
  • FI_BOR, FI_BAND, FI_BXOR (no floating point)
  • FI_ATOMIC_WRITE

Fetching (fi_fetch_atomic, etc.)

  • All of the basic operations as above
  • FI_ATOMIC_READ

Comparison (fi_compare_atomic, etc.)

  • FI_CSWAP
  • FI_MSWAP

FI_MSG

All FI_MSG operations are supported.

FI_RMA

All FI_RMA operations are supported.

FI_TAGGED

All FI_TAGGED operations are supported except fi_tinjectdata.

GNI EXTENSIONS

The GNI provider exposes low-level tuning parameters via a domain and endpoint fi_open_ops interface named FI_GNI_DOMAIN_OPS_1 and FI_GNI_EP_OPS_1. The flags parameter is currently ignored. The fi_open_ops function takes a struct fi_gni_ops_domain or a struct fi_gni_ops_ep parameter respectively and populates it with the following:

struct fi_gni_ops_domain {
	int (*set_val)(struct fid *fid, dom_ops_val_t t, void *val);
	int (*get_val)(struct fid *fid, dom_ops_val_t t, void *val);
	int (*flush_cache)(struct fid *fid);
};

struct fi_gni_ops_ep {
	int (*set_val)(struct fid *fid, dom_ops_val_t t, void *val);
	int (*get_val)(struct fid *fid, dom_ops_val_t t, void *val);
        size_t (*native_amo)(struct fid_ep *ep, const void *buf,
                             size_t count,void *desc,
                             fi_addr_t dest_addr, uint64_t addr,
                             uint64_t key, enum fi_datatype datatype,
                             enum gnix_fab_req_type req_type,
                             void *context);
};

The set_val function sets the value of a given parameter; the get_val function returns the current value. For FI_GNI_DOMAIN_OPS_1, the currently supported values are:

GNI_MSG_RENDEZVOUS_THRESHOLD
Threshold message size at which a rendezvous protocol is used for FI_MSG data transfers. The value is of type uint32_t.
GNI_RMA_RDMA_THRESHOLD
Threshold message size at which RDMA is used for FI_RMA data transfers. The value is of type uint32_t.
GNI_CONN_TABLE_INITIAL_SIZE
Initial size of the internal table data structure used to manage connections. The value is of type uint32_t.
GNI_CONN_TABLE_MAX_SIZE
Maximum size of the internal table data structure used to manage connections. The value is of type uint32_t.
GNI_CONN_TABLE_STEP_SIZE
Step size for increasing the size of the internal table data structure used to manage internal GNI connections. The value is of type uint32_t.
GNI_VC_ID_TABLE_CAPACITY
Size of the virtual channel (VC) table used for managing remote connections. The value is of type uint32_t.
GNI_MBOX_PAGE_SIZE
Page size for GNI SMSG mailbox allocations. The value is of type uint32_t.
GNI_MBOX_NUM_PER_SLAB
Number of GNI SMSG mailboxes per allocation slab. The value is of type uint32_t.
GNI_MBOX_MAX_CREDIT
Maximum number of credits per GNI SMSG mailbox. The value is of type uint32_t.
GNI_MBOX_MSG_MAX_SIZE
Maximum size of GNI SMSG messages. The value is of type uint32_t.
GNI_RX_CQ_SIZE
Recommended GNI receive CQ size. The value is of type uint32_t.
GNI_TX_CQ_SIZE
Recommended GNI transmit CQ size. The value is of type uint32_t.
GNI_MAX_RETRANSMITS
Maximum number of message retransmits before failure. The value is of type uint32_t.
GNI_MR_CACHE_LAZY_DEREG
Enable or disable lazy deregistration of memory. The value is of type int32_t.
GNI_MR_CACHE
Select the type of cache that the domain will use. Valid choices are the following: ‘internal’, ‘udreg’, or ‘none’. ‘internal’ refers to the GNI provider internal registration cache. ‘udreg’ refers to a user level dreg library based cache. Lastly, ‘none’ refers to device direct registration without a provider cache.
GNI_MR_HARD_REG_LIMIT
Maximum number of registrations. Applies only to the GNI provider cache. The value is of type int32_t (-1 for no limit).
GNI_MR_SOFT_REG_LIMIT
Soft cap on the registration limit. Applies only to the GNI provider cache. The value is of type int32_t (-1 for no limit).
GNI_MR_HARD_STALE_REG_LIMIT
Maximum number of stale registrations to be held in cache. This applies to the GNI provider cache and the udreg cache. The value is of type int32_t (-1 for no limit for the GNI provider cache and udreg cache values must be greater than 0).
GNI_MR_UDREG_LIMIT
Maximum number of registrations. Applies only to the udreg cache. The value is of type int32_t. The value must be greater than 0.
GNI_XPMEM_ENABLE
Enable or disable use of XPMEM for on node messages using the GNI provider internal rendezvous protocol. The value is of type bool.
GNI_DGRAM_PROGRESS_TIMEOUT
Controls timeout value in milliseconds for the control progress thread. The value is of type uint32_t.

The flush_cache function allows the user to flush any stale registration cache entries from the cache. This has the effect of removing registrations from the cache that have been deregistered with the provider, but still exist in case that they may be reused in the near future. Flushing the stale registrations forces hardware-level deregistration of the stale memory registrations and frees any memory related to those stale registrations. Only the provider-level registration struct is freed, not the user buffer associated with the registration. The parameter for flush_cache is a struct fid pointer to a fi_domain. The memory registration cache is tied to the domain, so issuing a flush_cache to the domain will flush the registration cache of the domain.

For FI_GNI_EP_OPS_1, the currently supported values are:
GNI_HASH_TAG_IMPL
Use a hashlist for the tag list implementation. The value is of type uint32_t.

The native_amo function allows the user to call GNI native atomics that are not implemented in the libfabric API. The parameters for native_amo are the same as the fi_atomic function but adds the following parameter:

enum gnix_fab_req_type req_type
The req_type’s supported with this call are GNIX_FAB_RQ_NAMO_AX (AND and XOR), and GNIX_FAB_RQ_NAMO_AX_S (AND and XOR 32 bit), GNIX_FAB_RQ_NAMO_FAX (Fetch AND and XOR) and GNIX_FAB_RQ_NAMO_FAX_S (Fetch AND and XOR 32 bit).

#NOTES

The default address format is FI_ADDR_GNI. This is the only address format used within the GNI provider for message passing. FI_ADDR_STR is always parsed and converted to FI_ADDR_GNI for use within the GNI provider.

FI_ADDR_STR is formatted as follows: gni;node;service;GNIX_AV_STR_ADDR_VERSION;device_addr;cdm_id;name_type;cm_nic_cdm_id;cookie;rx_ctx_cnt

The GNI provider sets the domain attribute cntr_cnt to the the CQ limit divided by 2.

The GNI provider sets the domain attribute cq_cnt to the CQ limit divided by 2.

The GNI provider sets the domain attribute ep_cnt to SIZE_MAX.

Completion queue events may report unknown source address information when using FI_SOURCE. The source address information will be reported in the err_data member of the struct fi_cq_err_entry populated by fi_cq_readerr. The err_data member will contain the source address information in the FI_ADDR_GNI address format. In order to populate the remote peer’s address vector with this mechanism, the application must call fi_cq_readerr to get the source address followed by fi_av_insert on the populated err_data member.

For FI_MULTI_RECV, the GNI provider generates a separate FI_MULTI_RECV CQ event once the receive buffer has been consumed. Also, owing to the out-or-order nature of the Cray network, the CQ events associated with individual messages arriving in the receive buffer may be generated out of order with respect to the offset into the buffer into which the messages were received.

KNOWN BUGS

The GNI provider currently treats the fi_shutdown() interface as a strictly local operation. That is, fi_shutdown() causes the local endpoint to be shut down, and a shutdown event to be generated on the local EQ. However, a connected remote peer endpoint is not notified of a call to fi_shutdown().

The GNI provider does not currently handle the case when FI_OPT_MULTI_RECV is set to 0 and will return -FI_EINVAL if an application attempts to set this value to zero.

SEE ALSO

fabric(7), fi_open_ops(3), fi_provider(7), fi_getinfo(3) fi_atomic(3)

For more information on uGNI, see Using the GNI and DMAPP APIs (S-2446-3103, Cray Inc.). For more information on the GNI provider, see An Implementation of OFI libfabric in Support of Multithreaded PGAS Solutions (PGAS ‘15).