1fi_efa(7) Libfabric v1.12.1 fi_efa(7)
2
3
4
6 fi_efa - The Amazon Elastic Fabric Adapter (EFA) Provider
7
9 The EFA provider supports the Elastic Fabric Adapter (EFA) device on
10 Amazon EC2. EFA provides reliable and unreliable datagram send/receive
11 with direct hardware access from userspace (OS bypass).
12
14 The following features are supported:
15
16 Endpoint types
17 The provider supports endpoint type FI_EP_DGRAM, and FI_EP_RDM
18 on a new Scalable (unordered) Reliable Datagram protocol (SRD).
19 SRD provides support for reliable datagrams and more complete
20 error handling than typically seen with other Reliable Datagram
21 (RD) implementations. The EFA provider provides segmentation,
22 reassembly of out-of-order packets to provide send-after-send
23 ordering guarantees to applications via its FI_EP_RDM endpoint.
24
25 RDM Endpoint capabilities
26 The following data transfer interfaces are supported via the
27 FI_EP_RDM endpoint: FI_MSG, FI_TAGGED, and FI_RMA. FI_SEND,
28 FI_RECV, FI_DIRECTED_RECV, FI_MULTI_RECV, and FI_SOURCE capabil‐
29 ities are supported. The endpoint provides send-after-send
30 guarantees for data operations. The FI_EP_RDM endpoint does not
31 have a maximum message size.
32
33 DGRAM Endpoint capabilities
34 The DGRAM endpoint only supports FI_MSG capability with a maxi‐
35 mum message size of the MTU of the underlying hardware (approxi‐
36 mately 8 KiB).
37
38 Address vectors
39 The provider supports FI_AV_TABLE and FI_AV_MAP address vector
40 types. FI_EVENT is unsupported.
41
42 Completion events
43 The provider supports FI_CQ_FORMAT_CONTEXT, FI_CQ_FORMAT_MSG,
44 and FI_CQ_FORMAT_DATA. FI_CQ_FORMAT_TAGGED is supported on the
45 RDM endpoint. Wait objects are not currently supported.
46
47 Modes The provider requires the use of FI_MSG_PREFIX when running over
48 the DGRAM endpoint, and requires FI_MR_LOCAL for all memory reg‐
49 istrations on the DGRAM endpoint.
50
51 Memory registration modes
52 The RDM endpoint does not require memory registration for send
53 and receive operations, i.e. it does not require FI_MR_LOCAL.
54 Applications may specify FI_MR_LOCAL in the MR mode flags in or‐
55 der to use descriptors provided by the application. The
56 FI_EP_DGRAM endpoint only supports FI_MR_LOCAL.
57
58 Progress
59 The RDM endpoint supports both FI_PROGRESS_AUTO and
60 FI_PROGRESS_MANUAL, with the default set to auto. However, re‐
61 ceive side data buffers are not modified outside of completion
62 processing routines. The DGRAM endpoint only supports
63 FI_PROGRESS_MANUAL.
64
65 Threading
66 The RDM endpoint supports FI_THREAD_SAFE, the DGRAM endpoint
67 supports FI_THREAD_DOMAIN, i.e. the provider is not thread safe
68 when using the DGRAM endpoint.
69
71 The DGRAM endpoint does not support FI_ATOMIC interfaces. For RMA op‐
72 erations, completion events for RMA targets (FI_RMA_EVENT) is not sup‐
73 ported. The DGRAM endpoint does not fully protect against resource
74 overruns, so resource management is disabled for this endpoint
75 (FI_RM_DISABLED).
76
77 No support for selective completions.
78
79 No support for counters for the DGRAM endpoint.
80
81 No support for inject.
82
84 FI_EFA_TX_SIZE
85 Maximum number of transmit operations before the provider re‐
86 turns -FI_EAGAIN. For only the RDM endpoint, this parameter
87 will cause transmit operations to be queued when this value is
88 set higher than the default and the transmit queue is full.
89
90 FI_EFA_RX_SIZE
91 Maximum number of receive operations before the provider returns
92 -FI_EAGAIN.
93
94 FI_EFA_TX_IOV_LIMIT
95 Maximum number of IOVs for a transmit operation.
96
97 FI_EFA_RX_IOV_LIMIT
98 Maximum number of IOVs for a receive operation.
99
101 These OFI runtime parameters apply only to the RDM endpoint.
102
103 FI_EFA_RX_WINDOW_SIZE
104 Maximum number of MTU-sized messages that can be in flight from
105 any single endpoint as part of long message data transfer.
106
107 FI_EFA_TX_QUEUE_SIZE
108 Depth of transmit queue opened with the NIC. This may not be
109 set to a value greater than what the NIC supports.
110
111 FI_EFA_RECVWIN_SIZE
112 Size of out of order reorder buffer (in messages). Messages re‐
113 ceived out of this window will result in an error.
114
115 FI_EFA_CQ_SIZE
116 Size of any cq created, in number of entries.
117
118 FI_EFA_MR_CACHE_ENABLE
119 Enables using the mr cache and in-line registration instead of a
120 bounce buffer for iov's larger than max_memcpy_size. Defaults
121 to true. When disabled, only uses a bounce buffer
122
123 FI_EFA_MR_MAX_CACHED_COUNT
124 Sets the maximum number of memory registrations that can be
125 cached at any time.
126
127 FI_EFA_MR_MAX_CACHED_SIZE
128 Sets the maximum amount of memory that cached memory registra‐
129 tions can hold onto at any time.
130
131 FI_EFA_MAX_MEMCPY_SIZE
132 Threshold size switch between using memory copy into a pre-reg‐
133 istered bounce buffer and memory registration on the user buf‐
134 fer.
135
136 FI_EFA_MTU_SIZE
137 Overrides the default MTU size of the device.
138
139 FI_EFA_RX_COPY_UNEXP
140 Enables the use of a separate pool of bounce-buffers to copy un‐
141 expected messages out of the pre-posted receive buffers.
142
143 FI_EFA_RX_COPY_OOO
144 Enables the use of a separate pool of bounce-buffers to copy
145 out-of-order RTS packets out of the pre-posted receive buffers.
146
147 FI_EFA_MAX_TIMEOUT
148 Maximum timeout (us) for backoff to a peer after a receiver not
149 ready error.
150
151 FI_EFA_TIMEOUT_INTERVAL
152 Time interval (us) for the base timeout to use for exponential
153 backoff to a peer after a receiver not ready error.
154
155 FI_EFA_ENABLE_SHM_TRANSFER
156 Enable SHM provider to provide the communication across all in‐
157 tra-node processes. SHM transfer will be disabled in the case
158 where ptrace protection is turned on. You can turn it off to
159 enable shm transfer.
160
161 FI_EFA_SHM_AV_SIZE
162 Defines the maximum number of entries in SHM provider's address
163 vector.
164
165 FI_EFA_SHM_MAX_MEDIUM_SIZE
166 Defines the switch point between small/medium message and large
167 message. The message larger than this switch point will be
168 transferred with large message protocol.
169
170 FI_EFA_INTER_MAX_MEDIUM_MESSAGE_SIZE
171 The maximum size for inter EFA messages to be sent by using
172 medium message protocol. Messages which can fit in one packet
173 will be sent as eager message. Messages whose sizes are smaller
174 than this value will be sent using medium message protocol.
175 Other messages will be sent using CTS based long message proto‐
176 col.
177
179 fabric(7), fi_provider(7), fi_getinfo(3)
180
182 OpenFabrics.
183
184
185
186Libfabric Programmer's Manual 2020-09-01 fi_efa(7)