0406c46bb4
fix indention of infiniband node
2018-07-16 11:00:15 +02:00
60f55ec178
improve naming of struct node_type function pointers ( closes #150 )
2018-07-16 11:00:15 +02:00
Dennis Potter
d9080fa1db
Cleaned up some obsolete code. Closes #176
2018-07-16 10:54:15 +02:00
Dennis Potter
d64d1e6f37
Implemented ib_reverse()
...
The only value that gets reversed is src- and dst address. Fixes #175 .
2018-07-16 10:35:48 +02:00
Dennis Potter
eb55dee920
Added fallback which sets mode to listening mode
...
Before, the node would throw an error as soon as it cannot connect to
the remote host. Now, it will throw a warning and switch to listening
mode (in which it will wait for another node to connect).
2018-07-15 16:37:52 +02:00
Dennis Potter
6444a9e337
Split configuration in an in and out part
2018-07-15 13:51:18 +02:00
Dennis Potter
6975854376
Solved problem with blocking thread
...
In the case that a node was already disconnected but not stopped,
rdma_cm_get_event always blocked and we coulnd't join the threads. This
is solved in this commit by registering SIGUSR1 to the CM event thread.
This bug originated in issue #152
2018-07-14 16:46:23 +02:00
Dennis Potter
95e2a7c20b
Added extra check for inline mode. Closes #164
...
The actual maximum size for inline mode is now returned to the user and
there is a check that inline_mode is either 0 or 1. Furthermore, this
commit includes a minor improvement in ib_write()
2018-07-14 15:20:24 +02:00
Dennis Potter
1d6ee5aec8
Node can determine if data should be send inline
...
The user can set the maximum size of the inline data and the node checks
if a sample can be send inline. This commit doesn't contain a info
message to the user about what the final max inline size will be. (The
HCA will probably change the value set by the user.)
2018-07-13 13:50:30 +02:00
Dennis Potter
b1b778f542
Added buffer to ib_write()
...
Now, ib_write() reads every cycle cnt values from the Completion Queue.
If it is not able to return them to the framework immediately, it
temporarily saves them on a stack.
ib_write() checks every cycle if the stack is non-empty and if it is
possible to return values from the stack to the framework.
2018-07-13 12:21:59 +02:00
Dennis Potter
4cd8fc7150
ib_write and ib_read handle memory in a way that the pool doesn't underrun now. Sending data inline is broken in this commit
2018-07-12 17:49:17 +02:00
Dennis Potter
72e627b327
Fixes #166 , all node interfaces are modified
...
The functions now look like this
int node_read(struct node *n, struct sample *smps[], unsigned cnt, unsigned *release);
int node_write(struct node *n, struct sample *smps[], unsigned cnt, unsigned *release);
This commit enables nodes to control how many samples will
be released by the framework through *release
2018-07-11 18:14:29 +02:00
Dennis Potter
29ff75fad3
Removed some obsolete code for completion queues
2018-07-08 15:32:28 +02:00
Dennis Potter
746fd2f694
Refactored ib_write in the same way as ib_read (described in #153 ). Merged separate completion queue polls to ib_write. Closes #167
2018-07-08 15:00:47 +02:00
Dennis Potter
ebb5446305
Refactored the way ib_read() handles the refence counts for the samples it uses. This is based on the algorithm described in issue #153
2018-07-08 14:05:48 +02:00
Dennis Potter
6150a36411
Changed all node_write() functions
2018-07-07 17:48:07 +02:00
Dennis Potter
4663f55e4b
Changed all node_read() functions to support a *cnt instead of cnt
2018-07-07 17:07:45 +02:00
Dennis Potter
06e7434d6c
Solved some state problems. This commit also solves #154 , which was caused by a non-terminated thread. (This thread will be removed in a later commit anyway
2018-07-07 15:34:07 +02:00
Dennis Potter
836adee4d6
Node is able to clean everything up and reconnect. Node can abort if it is in STARTED and in CONNECTED state
2018-07-07 14:36:23 +02:00
Dennis Potter
80da4801e1
Source and target successfully connect and node changes status from STATE_STARTED to STATE_CONNECTED in this commit. Next step will be to fix ib_stop and ib_disconnect to make the target able to accept new connections.
2018-07-07 13:08:08 +02:00
Dennis Potter
e2061e58fc
Events are now monitored in a separate thread. The segmentation faults we saw earlier were caused because we exited ib_start before we created a protection domain, which is used by memory_allocation
2018-07-07 12:49:22 +02:00
Dennis Potter
2bf122991c
Started to convert the RDMA_CM_EVENT loop to a separate thread and added a new state to the node. This commit is still broken
2018-07-05 18:26:32 +02:00
Dennis Potter
43dc305fde
Placed sanity checks to a separate function ib_check. Closes #151
2018-07-05 15:30:33 +02:00
Dennis Potter
f976ce5418
Added debug messages with different verbosity levels
2018-07-05 13:57:25 +02:00
Dennis Potter
781f405a48
Adhered to coding style
2018-07-04 19:04:08 +02:00
Dennis Potter
86363f06da
Added better handling for FLUSH_ERRs of receive work queue
2018-07-04 15:37:25 +02:00
Dennis Potter
8f52d167f5
Fixed wrong directory in include of ib.h
2018-07-04 15:26:22 +02:00
Dennis Potter
51519c06df
Implemented new memory system implementation in infiniband node
2018-07-04 15:15:24 +02:00
Dennis Potter
f3a5b81649
Merge branch 'infiniband-memory' into infiniband
2018-07-04 10:50:21 +02:00
Dennis Potter
0470ebda33
Fixed indentations
2018-07-03 18:01:49 +02:00
Dennis Potter
a825009457
Removed bugs with regard to the zero-copy implementation. This commit is able to send -r 5000 -l 50000 with villas pipe. With a higher -r, the source throws errors that there aren't any receive WRs, so probably pipe doesn't fill up the WRs fast enough
2018-07-03 17:39:06 +02:00
Dennis Potter
669d75a666
Basic implementation of zero-copying is done. The is still a problem with rread = hook_read_list. It doesn't return anything after the fourth read
2018-07-02 16:03:16 +02:00
bb70be0b2c
memory: refactored memory subsystem
2018-07-02 15:57:56 +02:00
Dennis Potter
c055010be1
ib_write() now takes data directly from the super pool instead of copying it. ib_read() still copies data and the code needs cleanup after zero-copy is implemented
2018-07-01 12:56:03 +02:00
Dennis Potter
c70dbe2263
Resolved memory_ib related compile and link errors
2018-06-30 18:20:30 +02:00
c01393bc82
infiniband: add new memory type
2018-06-29 17:37:10 +02:00
Dennis Potter
1e2d4158f0
Added possibility to add port dynamically, added some config warnings and added a WR refresh function to ib_read
2018-06-28 17:24:28 +02:00
Dennis Potter
95393eeb35
Forgot to acknowledge events which caused the rdma_cm_id to block.
2018-06-28 14:43:08 +02:00
Dennis Potter
5598f93582
Implemented disconnect function on source and target side. The cleanup function doesn't go through completely yet, probably because rdma_destroy_id blocks because not everything in the rdma_cm_id is destroyed yet.
2018-06-28 12:46:16 +02:00
Dennis Potter
1df18da3f2
Added Work Completion threads. Send Completion thread checks for errors, Receive Completion Queue is still empty.
2018-06-27 17:01:47 +02:00
Dennis Potter
7631ec98ed
Refactored the rdma_cm_id structure
2018-06-27 11:21:28 +02:00
Dennis Potter
9e5836001d
Node is able to send messages. A lot is hardcoded and this was only meant to be a first setup of a working node.
2018-06-27 10:37:46 +02:00
Dennis Potter
4fef5d67e5
Memory is registrated succesfully. Created framework to send data. At this moment, the receiver doesn't prepare Receive Work Requests. This is the first thing to fix after this commit
2018-06-25 18:21:44 +02:00
Dennis Potter
077355d8ba
Started to implement memory managemen and registration. Not yet working and in an early stage. Still some debugging and testing printfs
2018-06-24 13:02:04 +02:00
Dennis Potter
4f6c2543b5
Fixed bugs in connection. A source node is now able to connect to a target node.
2018-06-23 19:05:33 +02:00
Dennis Potter
a0f0410e73
This commit should be able to create a connection between two nodes.
...
This commit isn't checked for bugs and no memory management is done.
Furthermore, no poll threads are implemented.
2018-06-23 14:53:37 +02:00
Dennis Potter
4220ff8111
Implemented request to resolve address and built an rdma_cm_event framework
2018-06-22 13:02:41 +02:00
Dennis Potter
a1b7a9cda6
Initial commit with (empty) infiniband node
2018-06-21 11:56:28 +02:00
