Benchmark NVSHMEM 3.5.0 with EFA optimal patch

[HermesCui]

Hi, I ran the UCCl baseline benchmark using the latest NVSHMEM 3.5.0 version, which includes the EFA optimization patch. Here are the results from the baseline benchmark.
EP=16 DP=16

E	E/tok	tok	dim	CPU+Torch_total	CPU+Torch_lat	CPU+Torch_bw	CPU+Torch_comm	CPU_pack	CPU+NVSHMEM_total	CPU+NVSHMEM_lat	CPU+NVSHMEM_bw	CPU+NVSHMEM_comm	Dense_lat	Dense_bw	Sparse_bytes	Dense_bytes	CUDA+Torch_total	CUDA+Torch_lat	CUDA+Torch_bw	CUDA+Torch_comm	CUDA_pack	CUDA+NVSHMEM_total	CUDA+NVSHMEM_lat	CUDA+NVSHMEM_bw	CUDA+NVSHMEM_comm	Speedup_Torch	Speedup_NVSHMEM
256	8	128	7168	105417.6μs	30053.0μs	0.25GB/s	1659.9μs	75364.6μs	131503.7μs	56139.1μs	0.15GB/s	27780.5μs	1405.4μs	5.23GB/s	7,348,224	7,348,224	3956.4μs	1405.4μs	5.23GB/s	405.2μs	2551.0μs	28608.4μs	26057.4μs	0.82GB/s	25058.1μs	26.65x	4.60x
256	8	256	7168	199167.0μs	55347.9μs	0.27GB/s	525.5μs	143819.0μs	208128.5μs	64309.5μs	0.23GB/s	9547.2μs	2361.1μs	6.22GB/s	14,696,448	14,696,448	4421.7μs	2361.1μs	6.22GB/s	508.4μs	2060.5μs	5422.3μs	3361.8μs	4.37GB/s	1538.0μs	45.04x	38.38x
256	8	512	7168	396906.4μs	108975.4μs	0.27GB/s	797.7μs	287931.0μs	411028.2μs	123097.2μs	0.24GB/s	15033.0μs	4349.6μs	6.76GB/s	29,392,896	29,392,896	6412.1μs	4349.6μs	6.76GB/s	802.5μs	2062.5μs	8781.8μs	6719.3μs	4.37GB/s	3190.0μs	61.90x	46.80x
256	8	768	7168	592873.8μs	162597.7μs	0.27GB/s	1262.0μs	430276.1μs	615808.6μs	185532.6μs	0.24GB/s	24424.2μs	6164.4μs	7.15GB/s	44,089,344	44,089,344	8314.2μs	6164.4μs	7.15GB/s	1259.3μs	2149.8μs	11410.4μs	9260.6μs	4.76GB/s	4383.2μs	71.31x	53.97x
256	8	1024	7168	790688.1μs	216602.8μs	0.27GB/s	1530.3μs	574085.3μs	817837.3μs	243752.0μs	0.24GB/s	29121.8μs	8083.8μs	7.27GB/s	58,785,792	58,785,792	10276.3μs	8083.8μs	7.27GB/s	1532.1μs	2192.4μs	14813.5μs	12621.0μs	4.66GB/s	6088.6μs	76.94x	55.21x
256	8	2048	7168	1574841.9μs	429761.8μs	0.27GB/s	3006.6μs	1145080.0μs	1631731.0μs	486651.0μs	0.24GB/s	60564.4μs	15267.4μs	7.70GB/s	117,571,584	117,571,584	17592.7μs	15267.4μs	7.70GB/s	2899.0μs	2325.3μs	33208.4μs	30883.1μs	3.81GB/s	18545.3μs	89.52x	49.14x
256	8	4096	7168	3162618.5μs	863716.8μs	0.27GB/s	5975.8μs	2298901.8μs	3265733.0μs	966831.4μs	0.24GB/s	108283.3μs	30284.3μs	7.76GB/s	235,143,168	235,143,168	33907.1μs	30284.3μs	7.76GB/s	5710.9μs	3622.8μs	63082.3μs	59459.5μs	3.96GB/s	34931.3μs	93.27x	51.77x

[HermesCui]

This is result based on pplx kernel with NVSHMEM all2all

[YangZhou1997]

This seems more fitting for a PR with some README update on the results. @HermesCui can you open a PR and commit results to the repo?

[HermesCui]

Sure, I gonna open a pr to update ep/bench/baseline README