void Test(api::Context& ctx) {
for (size_t outer = 0; outer < outer_repeats_; ++outer) {
common::StatsTimerStopped t;
size_t dummy = +4915221495089;
t.Start();
for (size_t inner = 0; inner < inner_repeats_; ++inner) {
dummy = ctx.net.Broadcast(dummy);
}
t.Stop();
size_t n = ctx.num_workers();
size_t time = t.Microseconds();
// calculate maximum time.
time = ctx.net.AllReduce(time, common::maximum<size_t>());
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " datatype=" << "size_t"
<< " operation=" << "broadcast"
<< " workers=" << n
<< " inner_repeats=" << inner_repeats_
<< " time[us]=" << time
<< " time_per_op[us]="
<< static_cast<double>(time) / inner_repeats_
<< std::endl;
}
}
}
void Test(api::Context& ctx) {
common::StatsTimerStopped t;
// only work with first thread on this host.
if (ctx.local_worker_id() == 0)
{
mem::Manager mem_manager(nullptr, "Dispatcher");
group_ = &ctx.net.group();
std::unique_ptr<net::Dispatcher> dispatcher =
group_->ConstructDispatcher();
dispatcher_ = dispatcher.get();
t.Start();
for (size_t outer = 0; outer < outer_repeats_; ++outer)
{
rnd_ = std::default_random_engine(123456);
active_ = 0;
remaining_requests_ = num_requests_;
while (active_ < limit_active_ && remaining_requests_ > 0)
{
if (MaybeStartRequest()) {
++active_;
}
}
dispatcher_->Loop();
}
t.Stop();
// must clean up dispatcher prior to using group for other things.
}
size_t time = t.Microseconds();
// calculate maximum time.
time = ctx.net.AllReduce(time, common::maximum<size_t>());
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " operation=" << "rblocks"
<< " hosts=" << group_->num_hosts()
<< " requests=" << num_requests_
<< " block_size=" << block_size_
<< " limit_active=" << limit_active_
<< " time[us]=" << time
<< " time_per_op[us]="
<< static_cast<double>(time) / num_requests_
<< " total_bytes=" << block_size_ * num_requests_
<< " total_bandwidth[MiB/s]="
<< CalcMiBs(block_size_ * num_requests_, time)
<< std::endl;
}
}
void Test(api::Context& ctx) {
for (size_t outer = 0; outer < outer_repeats_; ++outer) {
common::StatsTimerStopped t;
size_t n = ctx.num_workers();
t.Start();
for (size_t inner = 0; inner < inner_repeats_; ++inner) {
// allreduce a different value in each iteration
size_t value = inner + ctx.my_rank();
value = ctx.net.AllReduce(value);
size_t expected = (n + inner) * ((n + inner) - 1) / 2 - inner * (inner - 1) / 2;
die_unequal(value, expected);
}
t.Stop();
size_t time = t.Microseconds();
// calculate maximum time.
time = ctx.net.AllReduce(time, common::maximum<size_t>());
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " datatype=" << "size_t"
<< " operation=" << "allreduce"
<< " workers=" << n
<< " inner_repeats=" << inner_repeats_
<< " time[us]=" << time
<< " time_per_op[us]="
<< static_cast<double>(time) / inner_repeats_
<< std::endl;
}
}
}
void ExperimentFull(
api::Context& ctx, const std::string& type_as_string) {
// transmit data to all workers.
auto stream = ctx.GetNewCatStream();
// write phase
StatsTimer<true> write_timer(true);
{
auto writers = stream->OpenWriters();
auto data = Generator<Type>(g_bytes);
while (data.HasNext()) {
Type value = data.Next();
for (size_t tgt = 0; tgt < ctx.num_workers(); ++tgt) {
writers[tgt](value);
}
}
}
write_timer.Stop();
// read phase
StatsTimer<true> read_timer(true);
{
auto reader = stream->OpenCatReader(true);
while (reader.HasNext()) {
reader.Next<Type>();
}
}
read_timer.Stop();
size_t read_microsecs = read_timer.Microseconds();
read_microsecs =
ctx.AllReduce(read_microsecs, common::maximum<size_t>());
size_t write_microsecs = write_timer.Microseconds();
write_microsecs =
ctx.AllReduce(write_microsecs, common::maximum<size_t>());
uint64_t host_volume = ctx.num_workers() * g_bytes;
uint64_t total_volume = ctx.num_workers() * ctx.num_workers() * g_bytes;
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " datatype=" << type_as_string
<< " size=" << g_bytes
<< " write_time=" << write_microsecs
<< " read_time=" << read_microsecs
<< " write_speed_MiBs="
<< (g_bytes / write_microsecs * 1000000 / 1024 / 1024)
<< " read_speed_MiBs="
<< (g_bytes / read_microsecs * 1000000 / 1024 / 1024)
<< " host_write_speed_MiBs="
<< (host_volume / write_microsecs * 1000000 / 1024 / 1024)
<< " host_read_speed_MiBs="
<< (host_volume / read_microsecs * 1000000 / 1024 / 1024)
<< " total_write_speed_MiBs="
<< (total_volume / write_microsecs * 1000000 / 1024 / 1024)
<< " total_read_speed_MiBs="
<< (total_volume / read_microsecs * 1000000 / 1024 / 1024)
<< std::endl;
}
}
void ExperimentAllPairs(
api::Context& ctx, const std::string& type_as_string) {
for (size_t src = 0; src < ctx.num_workers(); ++src) {
for (size_t tgt = 0; tgt < ctx.num_workers(); ++tgt) {
// transmit data from worker src -> tgt: only send data if we are
// tgt, but as tgt receive from all.
auto stream = ctx.GetNewCatStream();
// write phase
StatsTimer<true> write_timer(true);
{
auto writers = stream->OpenWriters();
if (ctx.my_rank() == src) {
auto data = Generator<Type>(g_bytes);
auto& writer = writers[tgt];
while (data.HasNext()) {
writer(data.Next());
}
}
}
write_timer.Stop();
// read phase
StatsTimer<true> read_timer(true);
{
auto reader = stream->OpenCatReader(true);
while (reader.HasNext()) {
reader.Next<Type>();
}
}
read_timer.Stop();
size_t read_microsecs = read_timer.Microseconds();
read_microsecs =
ctx.AllReduce(read_microsecs, common::maximum<size_t>());
size_t write_microsecs = write_timer.Microseconds();
write_microsecs =
ctx.AllReduce(write_microsecs, common::maximum<size_t>());
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " datatype=" << type_as_string
<< " size=" << g_bytes
<< " src=" << src << " tgt=" << tgt
<< " write_time=" << write_microsecs
<< " read_time=" << read_microsecs
<< " write_speed_MiBs="
<< (g_bytes / write_microsecs * 1000000 / 1024 / 1024)
<< " read_speed_MiBs="
<< (g_bytes / read_microsecs * 1000000 / 1024 / 1024)
<< std::endl;
}
}
}
}
void Bandwidth::Test(api::Context& ctx) {
// only work with first thread on this host.
if (ctx.local_worker_id() != 0) return;
net::Group& group = ctx.net.group();
bandwidth_ = AggMatrix(group.num_hosts());
// data block to send or receive
block_count_ = data_size_ / block_size_;
data_block_.resize(block_size_ / sizeof(size_t), 42u);
for (size_t outer_repeat = 0;
outer_repeat < outer_repeats_; ++outer_repeat) {
common::StatsTimerStopped timer;
timer.Start();
for (size_t inner_repeat = 0;
inner_repeat < inner_repeats_; inner_repeat++) {
// perform 1-factor ping pongs (without barriers)
for (size_t round = 0; round < group.OneFactorSize(); ++round) {
size_t peer = group.OneFactorPeer(round);
sLOG0 << "round" << round
<< "me" << ctx.host_rank() << "peer_id" << peer;
if (ctx.host_rank() < peer) {
Sender(ctx, peer, inner_repeat);
Receiver(ctx, peer);
}
else if (ctx.host_rank() > peer) {
Receiver(ctx, peer);
Sender(ctx, peer, inner_repeat);
}
else {
// not participating in this round
counter_ += 2 * block_count_;
}
}
}
timer.Stop();
size_t time = timer.Microseconds();
// calculate maximum time.
group.AllReduce(time, common::maximum<size_t>());
if (ctx.my_rank() == 0) {
std::cout
<< "RESULT"
<< " benchmark=" << benchmark
<< " hosts=" << ctx.num_hosts()
<< " outer_repeat=" << outer_repeat
<< " inner_repeats=" << inner_repeats_
<< " time[us]=" << time
<< " time_per_ping_pong[us]="
<< static_cast<double>(time) / static_cast<double>(counter_)
<< std::endl;
}
}
// reduce (add) matrix to root.
group.Reduce(bandwidth_);
// print matrix
if (ctx.my_rank() == 0)
PrintMatrix(bandwidth_);
}