diff --git a/HugeCTR/dlrm/README.md b/HugeCTR/dlrm/README.md
index 2a699497..9636915a 100644
--- a/HugeCTR/dlrm/README.md
+++ b/HugeCTR/dlrm/README.md
@@ -1,15 +1,17 @@
+
+
 # NVIDIA HugeCTR DLRM Benchmark Test 
+
 This folder holds NVIDIA HugeCTR DLRM Benchmark Test scripts, tools and reports.
 
 You can refer to [HugeCTR User Guide](https://github.com/NVIDIA/HugeCTR/blob/master/docs/hugectr_user_guide.md) for additional information.
 
-## folder structure
 ## Benchmark Test Cases
 
-This report summarized HugeCTR test on 1 nodes with 8 x Tesla V100 on Dec 2021
+This report summarized HugeCTR test on 1 nodes with 8 x Tesla V100 in Jan 2022
 
 ### Test Environment
-- 1 nodes with Tesla V100-SXM2-16GB x 8
+- 1 nodes with Tesla V100-SXM2-32GB x 8
 - InfiniBand 100 Gb/sec (4X EDR)， Mellanox Technologies MT27700 Family
 - Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz  ($ cat /proc/cpuinfo | grep name | cut -f2 -d: | uniq -c*)
 - Memory 384G ($ cat /proc/meminfo)
@@ -19,17 +21,17 @@ This report summarized HugeCTR test on 1 nodes with 8 x Tesla V100 on Dec 2021
 - `nvidia-smi topo -m`
 
 ```
-		GPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	mlx5_0	mlx5_1	CPU Affinity    NUMA Affinity
-GPU0	 X 		NV1	    NV2	    NV1	 	SYS		SYS		SYS		NV2		NODE	SYS		0-23,48-71		0
-GPU1	NV1	 	X 		NV1		NV2		SYS		SYS		NV2		SYS		NODE	SYS		0-23,48-71		0
-GPU2	NV2		NV1	 	X 		NV2		SYS		NV1		SYS		SYS		PIX		SYS		0-23,48-71		0
-GPU3	NV1		NV2		NV2		X 		NV1		SYS		SYS		SYS		PIX		SYS		0-23,48-71		0
-GPU4	SYS		SYS		SYS		NV1		X 		NV2		NV2		NV1		SYS		NODE	24-47,72-95		1
-GPU5	SYS		SYS		NV1		SYS		NV2	 	X 		NV1		NV2		SYS		NODE	24-47,72-95		1
-GPU6	SYS		NV2		SYS		SYS		NV2		NV1	 	X 		NV1		SYS		PIX		24-47,72-95		1
-GPU7	NV2		SYS		SYS		SYS		NV1		NV2		NV1	 	X 		SYS		PIX		24-47,72-95		1
-mlx5_0	NODE	NODE	PIX		PIX		SYS		SYS		SYS		SYS	 	X 		SYS		
-mlx5_1	SYS		SYS		SYS		SYS		NODE	NODE	PIX		PIX		SYS	 	X 		
+	GPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	mlx5_0	mlx5_1	CPU Affinity	NUMA Affinity
+GPU0	X	NV1	NV2	NV1	SYS	SYS	SYS	NV2	NODE	SYS	0-23,48-71	0
+GPU1	NV1	X	NV1	NV2	SYS	SYS	NV2	SYS	NODE	SYS	0-23,48-71	0
+GPU2	NV2	NV1	X	NV2	SYS	NV1	SYS	SYS	PIX	SYS	0-23,48-71	0
+GPU3	NV1	NV2	NV2	X	NV1	SYS	SYS	SYS	PIX	SYS	0-23,48-71	0
+GPU4	SYS	SYS	SYS	NV1	X 	NV2	NV2	NV1	SYS	NODE	24-47,72-95	1
+GPU5	SYS	SYS	NV1	SYS	NV2	X 	NV1	NV2	SYS	NODE	24-47,72-95	1
+GPU6	SYS	NV2	SYS	SYS	NV2	NV1	X 	NV1	SYS	PIX	24-47,72-95	1
+GPU7	NV2	SYS	SYS	SYS	NV1	NV2	NV1	X 	SYS	PIX	24-47,72-95	1
+mlx5_0	NODE	NODE	PIX	PIX	SYS	SYS	SYS	SYS	X	SYS		
+mlx5_1	SYS	SYS	SYS	SYS	NODE	NODE	PIX	PIX	SYS	X 		
 
 Legend:
 
@@ -66,9 +68,61 @@ command: bash dlrm.sh
 | decay_start                  | 0                              |
 | decay_steps                  | 1                              |
 | decay_power                  | 2                              |
-| end_lr',                     | 0                              |
+| end_lr                       | 0                              |
+
+### baseline runing log
+
+see baseline_log_info.csv
+
+### Test Result
+
+#### embedding size
+
+one GPU
+
+| gpu  | batch_size | embedding_vec_size | latency(ms) | memory_usage(MB) |
+| ---- | ---------- | ------------------ | ----------- | ---------------- |
+| n1g1 | 32768      | 2                  | 41.150      | 4,252            |
+| n1g1 | 32768      | 8                  | 41.396      | 5,286            |
+| n1g1 | 32768      | 32                 | 42.675      | 9,422            |
+| n1g1 | 32768      | 128                | 51.944      | 25,962           |
+
+eight GPUs
+
+| gpu  | batch_size | embedding_vec_size | latency(ms) | memory_usage(MB) |
+| ---- | ---------- | ------------------ | ----------- | ---------------- |
+| n1g8 | 32768      | 2                  | 66.358      | 2,108            |
+| n1g8 | 32768      | 8                  | 69.207      | 2,176            |
+| n1g8 | 32768      | 32                 | 64.588      | 2,448            |
+| n1g8 | 32768      | 128                | 201.402     | 3,536            |
+| n1g8 | 65536      | 128                | 357.724     | 5,406            |
+
+#### batch size
+
+one GPU
+
+| gpu  | batch_size | embedding_vec_size | latency(ms) | memory_usage(MB) |
+| ---- | ---------- | ------------------ | ----------- | ---------------- |
+| n1g1 | 16         | 128                | 0.508       | 17,884           |
+| n1g1 | 64         | 128                | 0.608       | 17,902           |
+| n1g1 | 256        | 128                | 0.957       | 17,940           |
+| n1g1 | 1024       | 128                | 2.159       | 18,134           |
+| n1g1 | 4096       | 128                | 7.060       | 18,896           |
+| n1g1 | 16384      | 128                | 26.455      | 21,928           |
+| n1g1 | 32768      | 128                | 51.935      | 25,962           |
+
+eight GPUs
+
+| gpu  | batch_size | embedding_vec_size | latency(ms) | memory_usage(MB) |
+| ---- | ---------- | ------------------ | ----------- | ---------------- |
+| n1g8 | 16         | 128                | 0.859       | 1,666            |
+| n1g8 | 64         | 128                | 1.136       | 1,674            |
+| n1g8 | 256        | 128                | 2.262       | 1,692            |
+| n1g8 | 1024       | 128                | 7.232       | 1,732            |
+| n1g8 | 4096       | 128                | 31.997      | 1,906            |
+| n1g8 | 16384      | 128                | 112.966     | 2,604            |
+| n1g8 | 32768      | 128                | 201.806     | 3,536            |
+| n1g8 | 65536      | 128                | 357.724     | 5,406            |
 
-### baseline 运行log
 
-见baseline_log_info.csv
 
diff --git a/HugeCTR/dlrm/dlrm.py b/HugeCTR/dlrm/dlrm.py
index cba254d8..cd68dce5 100644
--- a/HugeCTR/dlrm/dlrm.py
+++ b/HugeCTR/dlrm/dlrm.py
@@ -53,7 +53,7 @@ def DLRM(args):
     model.add(hugectr.DenseLayer(layer_type = hugectr.Layer_t.InnerProduct,
                                 bottom_names = ["relu2"],
                                 top_names = ["fc3"],
-                                num_output=128))
+                                num_output=args.embedding_vec_size))
     model.add(hugectr.DenseLayer(layer_type = hugectr.Layer_t.ReLU,
                                 bottom_names = ["fc3"],
                                 top_names = ["relu3"]))                              
@@ -126,7 +126,7 @@ def _print_args(args):
         print("=".ljust(66, "="))
         print(
             "Running {}: gpu_num_per_node = {}, num_nodes = {}.".format(
-                "HugeCTR-WDL", args.gpu_num_per_node, args.num_nodes
+                "HugeCTR-DLRM", args.gpu_num_per_node, args.num_nodes
             )
         )
         print("=".ljust(66, "="))
diff --git a/HugeCTR/dlrm/dlrm_bsz_test.sh b/HugeCTR/dlrm/dlrm_bsz_test.sh
new file mode 100644
index 00000000..d339a5af
--- /dev/null
+++ b/HugeCTR/dlrm/dlrm_bsz_test.sh
@@ -0,0 +1,24 @@
+max_iter=12000
+warmup_steps=1000
+lr=0.5
+for bsz in 16 64 256 1024 4096 16384 32768
+do
+    for ngpu in 1 8
+    do 
+        test_case=dlrm_test_n1g$ngpu}_bsz${bsz}
+        mem_usage_file=${test_case}.mem
+
+        python gpu_memory_usage.py 1> log/$mem_usage_file 2>&1 </dev/null &
+
+        python dlrm.py \
+            --gpu_num_per_node ${ngpu} \
+            --eval_batchs 70 \
+            --batch_size ${bsz} \
+            --learning_rate ${lr} \
+            --warmup_steps ${warmup_steps} \
+            --max_iter ${max_iter} \
+            --loss_print_every_n_iter 100 \
+            --embedding_vec_size 128 \
+            --eval_interval 100 | tee log/${test_case}.log
+    done
+done
diff --git a/HugeCTR/dlrm/dlrm_embedding_size_test.sh b/HugeCTR/dlrm/dlrm_embedding_size_test.sh
new file mode 100644
index 00000000..605c13b3
--- /dev/null
+++ b/HugeCTR/dlrm/dlrm_embedding_size_test.sh
@@ -0,0 +1,25 @@
+max_iter=12000
+warmup_steps=1000
+lr=0.5
+for embedding_vec_size in 2 8 32 128
+do
+        for ngpu in 1 8
+        do
+                test_case=dlrm_test_n1g${ngpu}_embsz${embedding_vec_size}
+                mem_usage_file=${test_case}.mem
+
+                python gpu_memory_usage.py 1> log/$mem_usage_file 2>&1 </dev/null &
+
+                python dlrm.py \
+                        --gpu_num_per_node ${ngpu} \
+                        --eval_batchs 70 \
+                        --batch_size 32768 \
+                        --learning_rate ${lr} \
+                        --warmup_steps ${warmup_steps} \
+                        --max_iter ${max_iter} \
+                        --loss_print_every_n_iter 100 \
+                        --embedding_vec_size ${embedding_vec_size} \
+                        --eval_interval 100 | tee log/${test_case}.log
+        done
+done
+
diff --git a/HugeCTR/dlrm/extract_hugectr_log.py b/HugeCTR/dlrm/extract_hugectr_log.py
new file mode 100644
index 00000000..7c35dbb2
--- /dev/null
+++ b/HugeCTR/dlrm/extract_hugectr_log.py
@@ -0,0 +1,115 @@
+import argparse
+import os
+import glob
+from statistics import median
+
+
+
+
+def write_line(f, lst, separator=',', start_end=False):
+    lst = ['', *lst, ''] if start_end else lst
+    f.write(separator.join(lst))
+    f.write('\n')
+
+
+def value_format(value):
+    if isinstance(value, float):
+        return '{:.3f}'.format(value)
+    elif isinstance(value, int):
+        return f'{value:,}'
+    else:
+        return str(value)
+
+
+def extract_mem_info(mem_file):
+    if not os.path.isfile(mem_file):
+        return 'NA'
+
+    with open(mem_file, 'r') as f:
+        for line in f.readlines():
+            ss = line.split(' ')
+            if len(ss) < 5:
+                continue
+            if ss[0] == 'max':
+                return int(float(ss[-1].strip()) / 1024 /1024)
+    return 'NA'
+
+
+def extract_info_from_file(log_file, start_iter):
+    '''
+    batch_size_per_device = 128
+    gpu_num_per_node = 8
+    num_nodes = 2
+    [HUGECTR][02:51:59][INFO][RANK0]: Iter: 100 Time(100 iters): 0.315066s Loss: 0.125152 lr:0.001000
+    [HUGECTR][02:51:59][INFO][RANK0]: Iter: 200 Time(100 iters): 0.213347s Loss: 0.106469 lr:0.001000
+    ...
+    [HUGECTR][02:52:01][INFO][RANK0]: Iter: 1100 Time(100 iters): 0.222373s Loss: 0.100451 lr:0.001000
+    max_iter = 1200
+    loss_print_every_n_iter = 100
+    gpu_num_per_node = 1
+    num_nodes = 1
+    '''
+    # extract info from file name
+    result_dict = {}
+    loss_print_every_n_iter = 0
+    with open(log_file, 'r') as f:
+        latencies = []
+        vocab_size = []
+        for line in f.readlines():
+            ss = line.split(' ')
+            if ss[0] in ['num_nodes', 'gpu_num_per_node', 'batch_size', 'embedding_vec_size']:
+                result_dict[ss[0]] = ss[2].strip()
+           # if ss[0] in ['workspace_size_per_gpu_in_mb']:
+                # result_dict['vocab_size'] = int((int(ss[2].strip()) * 1024 * 1024 / 4) // 1000000 * 1000000)
+            if ss[0] == 'loss_print_every_n_iter':
+                loss_print_every_n_iter = float(ss[2].strip())
+            elif len(ss) > 3 and ss[1] == 'Iter:' and '[INFO]' in ss[0]:
+                if int(ss[2].strip()) != start_iter:
+                    latencies.append(float(ss[5].strip()[:-1]))
+        if loss_print_every_n_iter > 0:
+            result_dict['latency(ms)'] = 1000 * sum(latencies) / len(latencies) / loss_print_every_n_iter
+    mem = extract_mem_info(log_file[:-3] + 'mem')
+    result_dict['memory_usage(MB)'] = mem
+    return result_dict
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="flags for HugeCTR WDL")
+    parser.add_argument("--benchmark_log_dir", type=str, required=True)
+    parser.add_argument("--start_iter", type=int, default=1000)
+    # parser.add_argument("--end_iter", type=int, default=1100)
+    args = parser.parse_args()
+
+    logs_list = sorted(glob.glob(os.path.join(args.benchmark_log_dir, "*.log")), key=os.path.getmtime)
+    #logs_list = sorted(logs_list)
+    chunk_list = {}
+    for log_file in logs_list:
+        test_result = extract_info_from_file(log_file, args.start_iter)
+        print(test_result)
+        json_file = os.path.basename(log_file)[:-4]
+        # print(json_file)
+        test_result['log_file'] = json_file
+        if json_file not in chunk_list.keys():
+            chunk_list[json_file] = []
+        chunk_list[json_file].append(test_result)
+    result_list = []
+    for log_name,chunk in chunk_list.items():
+        latency_list = []
+        for single_result in chunk:
+            latency_list.append(single_result['latency(ms)'])
+        tmp_chunk = chunk[0]
+        tmp_chunk['gpu'] = 'n{}g{}'.format(tmp_chunk['num_nodes'], tmp_chunk['gpu_num_per_node'])
+        tmp_chunk['latency(ms)'] = median(latency_list)
+        result_list.append(tmp_chunk)
+    #with open(os.path.join(args.benchmark_log_dir, 'latency_reprot.md'), 'w') as f:
+    report_file = args.benchmark_log_dir + '_latency_report.md'
+    with open(report_file, 'w') as f:
+        titles = ['log_file', 'gpu', 'batch_size', 'embedding_vec_size', 'latency(ms)', 'memory_usage(MB)']
+        write_line(f, titles, '|', True)
+        write_line(f, ['----' for _ in titles], '|', True)
+        for result in result_list:
+            if 'latency(ms)' not in result.keys():
+                print(result['log_file'], 'is not complete!')
+                continue
+            cells = [value_format(result[title]) for title in titles]
+            write_line(f, cells, '|', True)
diff --git a/HugeCTR/dlrm/gpu_memory_usage.py b/HugeCTR/dlrm/gpu_memory_usage.py
new file mode 100644
index 00000000..3f985b65
--- /dev/null
+++ b/HugeCTR/dlrm/gpu_memory_usage.py
@@ -0,0 +1,25 @@
+import time
+from pynvml import *
+
+nvmlInit()
+handle = nvmlDeviceGetHandleByIndex(0)
+running = True
+
+mem_threshold = 32*1024*1024
+state = 'init' #'Detecting'
+
+device0_max_used_mem = 0
+while running == True:
+    time.sleep(1)
+    info = nvmlDeviceGetMemoryInfo(handle)
+    if state == 'init':
+        if info.used > mem_threshold:
+            state = 'Detecting'
+    elif state == 'Detecting':
+        if info.used < mem_threshold:
+            running = False
+        else:
+            device0_max_used_mem = max(device0_max_used_mem, info.used)
+        
+nvmlShutdown()
+print('max device0 memory usage is:', device0_max_used_mem)