Skip to content

C
codes-dev

GitLab maintenance scheduled form Friday, 2021-06-18 5:00pm to Satursday, 2021-06-19 10:00pm CT - Services will be unavailable during this time.

Commit 9c4e2fad authored by Jonathan Jenkins's avatar Jonathan Jenkins
Browse files
Options

update modelnet interface, rc to impl local comm

parent a78fee52
Hide whitespace changes
Inline Side-by-side
Showing with 108 additions and 35 deletions
codes/model-net.h
View file @ 9c4e2fad
......@@ -20,6 +20,13 @@
#define CATEGORY_NAME_MAX 16
#define CATEGORY_MAX 12
// simple deprecation attribute hacking
#if defined(__GNUC__) || defined(__GNUG__) || defined(__clang__)
#define DEPRECATED(_msg) __attribute__((deprecated(_msg)))
#else
#define DEPRECATED(_msg)
#endif
/* HACK: there is currently no scheduling fidelity across multiple
* model_net_event calls. Hence, problems arise when some LP sends multiple
* messages as part of an event and expects FCFS ordering. A proper fix which
......@@ -27,14 +34,14 @@
* feasible for now (would basically have to redesign model-net), so expose
* explicit start-sequence and stop-sequence markers as a workaround
*/
extern int mn_in_seqence;
extern int mn_in_sequence;
extern tw_stime mn_msg_offset;
#define MN_START_SEQ() do {\
mn_in_seqence = 1; \
mn_in_sequence = 1; \
mn_msg_offset = 0.0; \
} while (0)
#define MN_END_SEQ() do {\
mn_in_seqence = 0;\
mn_in_sequence = 0;\
} while (0)
......@@ -65,6 +72,11 @@ enum msg_param_type {
MAX_MN_MSG_PARAM_TYPES
};
// return type for model_net_*event calls, to be passed into RC
// currently is just an int, but in the future may indicate whether lp-io was
// called etc.
typedef int model_net_event_return;
// network identifiers (both the config lp names and the model-net internal
// names)
extern char * model_net_lp_config_names[];
......@@ -175,7 +187,7 @@ void model_net_event_collective_rc(
* when calculating *both* sender and receiver modelnet LPs
*/
// first argument becomes the network ID
void model_net_event(
model_net_event_return model_net_event(
int net_id,
char const * category,
tw_lpid final_dest_lp,
......@@ -194,7 +206,7 @@ void model_net_event(
* annotation is not consulted here. The corresponding CODES map context is
* CODES_MCTX_GROUP_MODULO with the supplied annotation arguments.
*/
void model_net_event_annotated(
model_net_event_return model_net_event_annotated(
int net_id,
char const * annotation,
char const * category,
......@@ -212,7 +224,7 @@ void model_net_event_annotated(
* This variant uses CODES map contexts to calculate the sender and receiver
* modelnet LPs
*/
void model_net_event_mctx(
model_net_event_return model_net_event_mctx(
int net_id,
struct codes_mctx const * send_map_ctx,
struct codes_mctx const * recv_map_ctx,
......@@ -257,11 +269,18 @@ int model_net_get_msg_sz(int net_id);
/* NOTE: we may end up needing additoinal arguments here to track state for
* reverse computation; add as needed
*/
DEPRECATED("use model_net_event_rc2 instead, invalid RNG rollback can occur otherwise")
void model_net_event_rc(
int net_id,
tw_lp *sender,
uint64_t message_size);
/* This function replaces model_net_event_rc, and will replace the name later
* on. The num_rng_calls argument is the return value of model_net_*event*
* calls */
void model_net_event_rc2(
tw_lp *sender,
model_net_event_return const * ret);
/* Issue a 'pull' from the memory of the destination LP, without
* requiring the destination LP to do event processing. This is meant as a
......@@ -275,7 +294,7 @@ void model_net_event_rc(
* - self_event_size, self_event are applied at the requester upon receipt of
* the payload from the dest
*/
void model_net_pull_event(
model_net_event_return model_net_pull_event(
int net_id,
char const *category,
tw_lpid final_dest_lp,
......@@ -284,7 +303,7 @@ void model_net_pull_event(
int self_event_size,
void const *self_event,
tw_lp *sender);
void model_net_pull_event_annotated(
model_net_event_return model_net_pull_event_annotated(
int net_id,
char const * annotation,
char const *category,
......@@ -294,7 +313,7 @@ void model_net_pull_event_annotated(
int self_event_size,
void const *self_event,
tw_lp *sender);
void model_net_pull_event_mctx(
model_net_event_return model_net_pull_event_mctx(
int net_id,
struct codes_mctx const * send_map_ctx,
struct codes_mctx const * recv_map_ctx,
......@@ -305,6 +324,8 @@ void model_net_pull_event_mctx(
int self_event_size,
void const *self_event,
tw_lp *sender);
DEPRECATED("use model_net_event_rc2 instead, invalid RNG rollback can occur otherwise")
void model_net_pull_event_rc(
int net_id,
tw_lp *sender);
......
src/models/networks/model-net/model-net.c
View file @ 9c4e2fad
......@@ -43,7 +43,7 @@ struct model_net_method* method_array[] = {
#undef X
// counter and offset for the MN_START_SEQ / MN_END_SEQ macros
int mn_in_seqence = 0;
int mn_in_sequence = 0;
tw_stime mn_msg_offset = 0.0;
// message parameters for use via model_net_set_msg_param
......@@ -221,7 +221,42 @@ struct mn_stats* model_net_find_stats(char const * category, mn_stats mn_stats_a
return(&mn_stats_array[i]);
}
static void model_net_event_impl_base(
static model_net_event_return model_net_noop_event(
tw_lpid final_dest_lp,
int is_pull,
tw_stime offset,
int remote_event_size,
void const * remote_event,
int self_event_size,
void const * self_event,
tw_lp *sender)
{
model_net_event_return num_rng_calls = 0;
tw_stime poffset = mn_in_sequence ? mn_msg_offset : 0.0;
if (self_event_size && self_event != NULL) {
poffset += codes_local_latency(sender);
num_rng_calls++;
tw_event *e = tw_event_new(sender->gid, poffset+offset, sender);
memcpy(tw_event_data(e), self_event, self_event_size);
tw_event_send(e);
}
if (remote_event_size && remote_event != NULL) {
poffset += codes_local_latency(sender);
num_rng_calls++;
tw_event *e = tw_event_new(final_dest_lp, poffset+offset, sender);
memcpy(tw_event_data(e), remote_event, remote_event_size);
tw_event_send(e);
}
if (mn_in_sequence)
mn_msg_offset += poffset;
return num_rng_calls;
}
static model_net_event_return model_net_event_impl_base(
int net_id,
struct codes_mctx const * send_map_ctx,
struct codes_mctx const * recv_map_ctx,
......@@ -242,7 +277,7 @@ static void model_net_event_impl_base(
"%d but ROSS is configured for events of size %zd\n",
remote_event_size+self_event_size+sizeof(model_net_wrap_msg),
g_tw_msg_sz);
return;
return -1;
}
tw_lpid src_mn_lp = model_net_find_local_device_mctx(net_id, send_map_ctx,
......@@ -250,13 +285,19 @@ static void model_net_event_impl_base(
tw_lpid dest_mn_lp = model_net_find_local_device_mctx(net_id, recv_map_ctx,
final_dest_lp);
if (src_mn_lp == dest_mn_lp)
return model_net_noop_event(final_dest_lp, is_pull, offset,
remote_event_size, remote_event, self_event_size, self_event,
sender);
tw_stime poffset = codes_local_latency(sender);
if (mn_in_seqence){
if (mn_in_sequence){
tw_stime tmp = mn_msg_offset;
mn_msg_offset += poffset;
poffset += tmp;
}
tw_event *e = codes_event_new(src_mn_lp, poffset+offset, sender);
tw_event *e = tw_event_new(src_mn_lp, poffset+offset, sender);
model_net_wrap_msg *m = tw_event_data(e);
msg_set_header(model_net_base_magic, MN_BASE_NEW_MSG, sender->gid, &m->h);
......@@ -276,7 +317,7 @@ static void model_net_event_impl_base(
// this is an outgoing message
m->msg.m_base.is_from_remote = 0;
// set the msg-specific params
if (is_msg_params_set[MN_SCHED_PARAM_PRIO])
m->msg.m_base.sched_params = sched_params;
......@@ -297,12 +338,14 @@ static void model_net_event_impl_base(
//print_base(m);
tw_event_send(e);
return 1;
}
static void model_net_event_impl_base_rc(tw_lp *sender){
codes_local_latency_reverse(sender);
}
void model_net_event(
model_net_event_return model_net_event(
int net_id,
char const * category,
tw_lpid final_dest_lp,
......@@ -314,13 +357,13 @@ void model_net_event(
void const * self_event,
tw_lp *sender)
{
model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT, CODES_MCTX_DEFAULT,
category, final_dest_lp, message_size, 0, offset,
remote_event_size, remote_event, self_event_size, self_event,
sender);
return model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT,
CODES_MCTX_DEFAULT, category, final_dest_lp, message_size, 0,
offset, remote_event_size, remote_event, self_event_size,
self_event, sender);
}
void model_net_event_annotated(
model_net_event_return model_net_event_annotated(
int net_id,
char const * annotation,
char const * category,
......@@ -333,12 +376,12 @@ void model_net_event_annotated(
void const * self_event,
tw_lp *sender){
struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, true);
model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
message_size, 0, offset, remote_event_size, remote_event,
self_event_size, self_event, sender);
}
void model_net_event_mctx(
model_net_event_return model_net_event_mctx(
int net_id,
struct codes_mctx const * send_map_ctx,
struct codes_mctx const * recv_map_ctx,
......@@ -351,12 +394,13 @@ void model_net_event_mctx(
int self_event_size,
void const * self_event,
tw_lp *sender){
model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx, category,
final_dest_lp, message_size, 0, offset, remote_event_size,
remote_event, self_event_size, self_event, sender);
return model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx,
category, final_dest_lp, message_size, 0, offset,
remote_event_size, remote_event, self_event_size, self_event,
sender);
}
void model_net_pull_event(
model_net_event_return model_net_pull_event(
int net_id,
char const *category,
tw_lpid final_dest_lp,
......@@ -367,12 +411,12 @@ void model_net_pull_event(
tw_lp *sender){
/* NOTE: for a pull, we are filling the *remote* event - it will be remote
* from the destination's POV */
model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT, CODES_MCTX_DEFAULT,
category, final_dest_lp, message_size, 1, offset, self_event_size,
self_event, 0, NULL, sender);
return model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT,
CODES_MCTX_DEFAULT, category, final_dest_lp, message_size, 1,
offset, self_event_size, self_event, 0, NULL, sender);
}
void model_net_pull_event_annotated(
model_net_event_return model_net_pull_event_annotated(
int net_id,
char const * annotation,
char const *category,
......@@ -385,12 +429,12 @@ void model_net_pull_event_annotated(
/* NOTE: for a pull, we are filling the *remote* event - it will be remote
* from the destination's POV */
struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, true);
model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
message_size, 1, offset, self_event_size, self_event, 0, NULL,
sender);
}
void model_net_pull_event_mctx(
model_net_event_return model_net_pull_event_mctx(
int net_id,
struct codes_mctx const * send_map_ctx,
struct codes_mctx const * recv_map_ctx,
......@@ -403,11 +447,19 @@ void model_net_pull_event_mctx(
tw_lp *sender){
/* NOTE: for a pull, we are filling the *remote* event - it will be remote
* from the destination's POV */
model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx, category,
final_dest_lp, message_size, 1, offset, self_event_size,
return model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx,
category, final_dest_lp, message_size, 1, offset, self_event_size,
self_event, 0, NULL, sender);
}
void model_net_event_rc2(
tw_lp *sender,
model_net_event_return const * ret)
{
for (int i = 0; i < *ret; i++)
codes_local_latency_reverse(sender);
}
void model_net_event_rc(
int net_id,
tw_lp *sender,
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment