fabtests/benchmark_shared: bandwidth_rma: force sync after each warmup iteration

Modify bandwidth_rma() to synchronize after every warmup iteration
instead of only at window boundaries. This ensures proper warmup of
the server->client connection path for unidirectinal mode

The writedata test asymmetry (client sends many writes, server replies
with single send per window) means only the client->server path gets
warmed up during normal iterations. Forcing sync during warmup ensures
both directions are properly established before timing begins.

Also increase default warmup iterations from 10 to 16.

Signed-off-by: Alexey Novikov <[email protected]>

Author: alekswn

fabtests/benchmarks/benchmark_shared.c

@@ -767,7 +767,7 @@ int bandwidth_rma(enum ft_rma_opcodes rma_op, struct fi_rma_iov *remote)
 		if (ret)
 			return ret;
 
-		if (++j == opts.window_size) {
+		if (++j == opts.window_size || i < opts.warmup_iterations) {
 			ret = bw_rma_comp(rma_op, j);
 			if (ret)
 				return ret;

fabtests/include/shared.h

@@ -314,7 +314,7 @@ extern char default_port[8];
 #define INIT_OPTS (struct ft_opts) \
 	{	.options = FT_OPT_RX_CQ | FT_OPT_TX_CQ, \
 		.iterations = 1000, \
-		.warmup_iterations = 10, \
+		.warmup_iterations = 16, \
 		.num_connections = 1, \
 		.transfer_size = 1024, \
 		.window_size = 64, \