diff --git a/tools/urma_admin/admin_parameters.c b/tools/urma_admin/admin_parameters.c
index cd1ac6fdcbff6482963e38eb7fff135fddfda426..c01506a109392a1f930189c4e509853aa34666ba 100644
--- a/tools/urma_admin/admin_parameters.c
+++ b/tools/urma_admin/admin_parameters.c
@@ -328,6 +328,7 @@ int admin_parse_args(int argc, char *argv[], tool_config_t *cfg)
                 ret = admin_str_to_u32(optarg, &cfg->key.key_cnt);
                 break;
             case 'h':
+                cfg->help = true;
                 usage(argv[0]);
                 return 0;
             case 'd':
diff --git a/tools/urma_admin/admin_parameters.h b/tools/urma_admin/admin_parameters.h
index 0117427d8cc5c85314751d2c6fb2c25336aded29..283d16a4f867a4ac2dd27569329a58ed09b3f46e 100644
--- a/tools/urma_admin/admin_parameters.h
+++ b/tools/urma_admin/admin_parameters.h
@@ -260,6 +260,7 @@ typedef struct tool_config {
     tool_cmd_type_t cmd;
     bool specify_device;
     bool whole_info;
+    bool help;
     char dev_name[URMA_ADMIN_MAX_DEV_NAME];       /* ubep device name */
     urma_eid_t eid;
     bool dynamic_eid_mode;
diff --git a/tools/urma_admin/urma_admin.c b/tools/urma_admin/urma_admin.c
index 41b0b74b60e5b27607deca2d8d77f08afdf8a133..8f1a7719987f25e4f3bb4170923d96e53457db85 100644
--- a/tools/urma_admin/urma_admin.c
+++ b/tools/urma_admin/urma_admin.c
@@ -496,7 +496,7 @@ static void admin_log_cmd(int argc, char *argv[], int ret)
 int main(int argc, char *argv[])
 {
     int ret;
-    tool_config_t tool_cfg;
+    tool_config_t tool_cfg = {0};
 
     if (admin_check_cmd_len(argc, argv) != 0) {
         (void)printf("user: %s, cmd len out of range.\n", getlogin());
@@ -509,6 +509,10 @@ int main(int argc, char *argv[])
         admin_log_cmd(argc, argv, ret);
         return ret;
     }
+    if (tool_cfg.help) {
+        /* Do not execute other operations for --help parameter */
+        return 0;
+    }
     if (tool_cfg.cmd == TOOL_CMD_NUM) {
         admin_log_cmd(argc, argv, ret);
         return 0;
diff --git a/tools/urma_perftest/perftest_parameters.c b/tools/urma_perftest/perftest_parameters.c
index 709dc02d0005c0c3ded7e694f7e5a9635116c3f2..043b4e1e819965f278c0c641b909767a3d43a4eb 100644
--- a/tools/urma_perftest/perftest_parameters.c
+++ b/tools/urma_perftest/perftest_parameters.c
@@ -538,16 +538,20 @@ int perftest_parse_args(int argc, char *argv[], perftest_config_t *cfg)
                 break;
             case 'j':
                 if (ub_str_to_bool(optarg, &cfg->share_jfr) != 0) {
-                    (void)fprintf(stderr, "Invalid parameter :%s.\n", optarg);
+                    (void)fprintf(stderr, "Invalid parameter(share_jfr).\n");
                     return -1;
                 }
                 break;
             case 'J':
-                if (cfg->type != PERFTEST_BW) {
+                (void)ub_str_to_u32(optarg, &cfg->jettys);
+                if (cfg->jettys == 0) {
+                    (void)fprintf(stderr, "Invalid parameter(jettys).\n");
+                    return -1;
+                }
+                if (cfg->type != PERFTEST_BW && cfg->jettys > 1) {
                     (void)fprintf(stderr, "Multiple jettys only available on band width tests.\n");
                     return -1;
                 }
-                (void)ub_str_to_u32(optarg, &cfg->jettys);
                 break;
             case 'K':
                 (void)ub_str_to_u32(optarg, &cfg->token_policy);
diff --git a/tools/urma_perftest/perftest_parameters.h b/tools/urma_perftest/perftest_parameters.h
index e594dd4234aeb6fb94fedb1c6a38d7669518a1cd..aefc187058c488aaaa31ff604c9d9d7b22a962af 100644
--- a/tools/urma_perftest/perftest_parameters.h
+++ b/tools/urma_perftest/perftest_parameters.h
@@ -152,7 +152,8 @@ enum perftest_opts {
     PERFTEST_OPT_ENABLE_CREDIT,
     PERFTEST_OPT_CREDIT_THRESHOLD,
     PERFTEST_OPT_CREDIT_NOTIFY_CNT,
-    PERFTEST_OPT_JETTYS_PRE_JFR
+    PERFTEST_OPT_JETTYS_PRE_JFR,
+    PERFTEST_OPT_SEG_PRE_JETTY,
 };
 
 typedef enum perftest_rate_limiter_units {
@@ -231,6 +232,7 @@ typedef struct perftest_config {
 
     /* valid for server side of send_bw using iteraion type */
     int check_alive_exited;
+    bool seg_pre_jetty;
 } perftest_config_t;
 
 typedef struct perftest_value_range {
diff --git a/tools/urma_perftest/perftest_resources.c b/tools/urma_perftest/perftest_resources.c
index ed21f9c363af126dcdeac9617f2f43b7eaea140a..db8d2ab6801181615900d0e9e530f80e31eacd60 100644
--- a/tools/urma_perftest/perftest_resources.c
+++ b/tools/urma_perftest/perftest_resources.c
@@ -537,26 +537,35 @@ static void destroy_duplex_jettys(perftest_context_t *ctx, perftest_config_t *cf
 
 static inline void unregister_seg(perftest_context_t *ctx, const perftest_config_t *cfg, const int idx)
 {
+    uint32_t seg_num = (cfg->seg_pre_jetty == false) ? 1 : cfg->jettys;
     for (int k = 0; k < idx; k++) {
-        (void)urma_unregister_seg(ctx->local_tseg[k]);
+        if (k < seg_num) {
+            (void)urma_unregister_seg(ctx->local_tseg[k]);
+        }
     }
     free(ctx->local_tseg);
     ctx->local_tseg = NULL;
 }
 
-static inline void free_token_id(perftest_context_t *ctx, const int idx)
+static inline void free_token_id(perftest_context_t *ctx, const perftest_config_t *cfg, const int idx)
 {
+    uint32_t seg_num = (cfg->seg_pre_jetty == false) ? 1 : cfg->jettys;
     for (int k = 0; k < idx; k++) {
-        (void)urma_free_token_id(ctx->token_id[k]);
+        if (k < seg_num) {
+            (void)urma_free_token_id(ctx->token_id[k]);
+        }
     }
     free(ctx->token_id);
     ctx->token_id = NULL;
 }
 
-static inline void free_memory(perftest_context_t *ctx, const int idx)
+static inline void free_memory(perftest_context_t *ctx, const perftest_config_t *cfg, const int idx)
 {
+    uint32_t seg_num = (cfg->seg_pre_jetty == false) ? 1 : cfg->jettys;
     for (int j = 0; j < idx; j++) {
-        free(ctx->local_buf[j]);
+        if (j < seg_num) {
+            free(ctx->local_buf[j]);
+        }
     }
     free(ctx->local_buf);
     ctx->local_buf = NULL;
@@ -570,17 +579,23 @@ static int register_mem(perftest_context_t *ctx, perftest_config_t *cfg)
         return -ENOMEM;
     }
 
+    uint32_t seg_num = (cfg->seg_pre_jetty == false) ? 1 : cfg->jettys;
+
     ctx->page_size = cfg->page_size;
 
     // holds the size of maximum between cfg->size and page_size, aligned to cache line.
     uint64_t max_size = MAX(cfg->size, ctx->page_size);
     ctx->buf_size = PERFTEST_ALIGN_CACHELINE(max_size, cfg->cache_line_size);
     // Buff is divided into two parts, one for recv and the other for send
-    ctx->buf_len = ctx->buf_size * PERFTEST_BUF_NUM;
+    ctx->buf_len = ctx->buf_size * PERFTEST_BUF_NUM *
+        ((cfg->seg_pre_jetty == true) ? 1 : cfg->jettys);
 
     for (i = 0; i < (int)cfg->jettys; i++) {
-        // Buff is divided into two parts, one for recv and the other for send
-        ctx->local_buf[i] = memalign(ctx->page_size, ctx->buf_len);
+        if (i < seg_num) {
+            ctx->local_buf[i] = memalign(ctx->page_size, ctx->buf_len);
+        } else {
+            ctx->local_buf[i] = ctx->local_buf[0];
+        }
         if (ctx->local_buf[i] == NULL) {
             (void)fprintf(stderr, "Failed to memalign local buff, loop:%d!\n", i);
             goto free_memory;
@@ -593,7 +608,11 @@ static int register_mem(perftest_context_t *ctx, perftest_config_t *cfg)
             goto free_memory;
         }
         for (k = 0; k < (int)cfg->jettys; k++) {
-            ctx->token_id[k] = urma_alloc_token_id(ctx->urma_ctx);
+            if (k < seg_num) {
+                ctx->token_id[k] = urma_alloc_token_id(ctx->urma_ctx);
+            } else {
+                ctx->token_id[k] = ctx->token_id[0];
+            }
             if (ctx->token_id[k] == NULL) {
                 (void)fprintf(stderr, "Failed to alloc token id, loop:%d!\n", k);
                 goto free_token_id;
@@ -615,18 +634,22 @@ static int register_mem(perftest_context_t *ctx, perftest_config_t *cfg)
     };
     urma_seg_cfg_t seg_cfg = {
         .va = 0,
-        .len = ctx->buf_size  * PERFTEST_BUF_NUM,
+        .len = ctx->buf_len,
         .token_value = g_perftest_token,
         .flag = flag,
         .user_ctx = (uintptr_t)NULL,
         .iova = 0
     };
     for (j = 0; j < (int)cfg->jettys; j++) {
-        seg_cfg.va = (uint64_t)ctx->local_buf[j];
-        if (ctx->urma_ctx->dev->type == URMA_TRANSPORT_UB) {
-            seg_cfg.token_id = ctx->token_id[j];
+        if (j < seg_num) {
+            seg_cfg.va = (uint64_t)ctx->local_buf[j];
+            if (ctx->urma_ctx->dev->type == URMA_TRANSPORT_UB) {
+                seg_cfg.token_id = ctx->token_id[j];
+            }
+            ctx->local_tseg[j] = urma_register_seg(ctx->urma_ctx, &seg_cfg);
+        } else {
+            ctx->local_tseg[j] = ctx->local_tseg[0];
         }
-        ctx->local_tseg[j] = urma_register_seg(ctx->urma_ctx, &seg_cfg);
         if (ctx->local_tseg[j] == NULL) {
             (void)fprintf(stderr, "Failed to register seg, loop:%d!\n", j);
             goto unregister_seg;
@@ -639,10 +662,10 @@ unregister_seg:
     unregister_seg(ctx, cfg, j);
 free_token_id:
     if (ctx->urma_ctx->dev->type == URMA_TRANSPORT_UB) {
-        free_token_id(ctx, k);
+        free_token_id(ctx, cfg, k);
     }
 free_memory:
-    free_memory(ctx, i);
+    free_memory(ctx, cfg, i);
     return -1;
 }
 
@@ -650,9 +673,9 @@ static void unregister_mem(perftest_context_t *ctx, const perftest_config_t *cfg
 {
     unregister_seg(ctx, cfg, (int)ctx->jetty_num);
     if (ctx->urma_ctx->dev->type == URMA_TRANSPORT_UB) {
-        free_token_id(ctx, (int)ctx->jetty_num);
+        free_token_id(ctx, cfg, (int)ctx->jetty_num);
     }
-    free_memory(ctx, (int)ctx->jetty_num);
+    free_memory(ctx, cfg, (int)ctx->jetty_num);
 }
 
 static inline void free_remote_seg(perftest_context_t *ctx, const int idx)
diff --git a/tools/urma_perftest/perftest_run_test.c b/tools/urma_perftest/perftest_run_test.c
index 986097c4e131b12525a98da557706ac70cead0b0..2d14eb80853b339d4b5fb021fe0fc6fbf32ad4db 100644
--- a/tools/urma_perftest/perftest_run_test.c
+++ b/tools/urma_perftest/perftest_run_test.c
@@ -387,7 +387,8 @@ static int run_once_send_lat(perftest_context_t *ctx, perftest_config_t *cfg)
                     }
 
                     if (cr.status != URMA_CR_SUCCESS) {
-                        (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu", (int)cr.status, scnt, rcnt);
+                        (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu.\n",
+                            (int)cr.status, scnt, rcnt);
                         return -1;
                     }
 
@@ -456,7 +457,7 @@ static int run_once_send_lat(perftest_context_t *ctx, perftest_config_t *cfg)
                     return -1;
                 }
                 if (cr.status != URMA_CR_SUCCESS) {
-                    (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu", (int)cr.status, scnt, rcnt);
+                    (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu.\n", (int)cr.status, scnt, rcnt);
                     return -1;
                 }
 
@@ -778,8 +779,12 @@ static void init_read_jfs_wr_sg(urma_jfs_wr_t *wr, perftest_context_t *ctx, perf
     uint32_t i, uint32_t j)
 {
     run_test_ctx_t *run_ctx = &ctx->run_ctx;
-    uint64_t lva = (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
-    uint64_t rva = (uint64_t)ctx->remote_seg[i]->ubva.va;
+    uint64_t lva = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->local_buf[0] + (cfg->jettys + i) * ctx->buf_size :
+        (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
+    uint64_t rva = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->remote_seg[0]->ubva.va + i * ctx->buf_size :
+        (uint64_t)ctx->remote_seg[i]->ubva.va;
     urma_sge_t *local_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR + 1];
     urma_sge_t *remote_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR];
 
@@ -818,8 +823,13 @@ static void init_write_jfs_wr_sg(urma_jfs_wr_t *wr, perftest_context_t *ctx, per
     uint32_t i, uint32_t j)
 {
     run_test_ctx_t *run_ctx = &ctx->run_ctx;
-    uint64_t lva = (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
-    uint64_t rva = (uint64_t)ctx->remote_seg[i]->ubva.va;
+
+    uint64_t lva = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->local_buf[0] + (cfg->jettys + i) * ctx->buf_size :
+        (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
+    uint64_t rva = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->remote_seg[0]->ubva.va + i * ctx->buf_size :
+        (uint64_t)ctx->remote_seg[i]->ubva.va;
     urma_sge_t *local_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR + 1];
     urma_sge_t *remote_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR];
 
@@ -859,7 +869,9 @@ static void init_send_jfs_wr_sg(urma_jfs_wr_t *wr, perftest_context_t *ctx, perf
     uint32_t i, uint32_t j)
 {
     run_test_ctx_t *run_ctx = &ctx->run_ctx;
-    uint64_t lva = (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
+    uint64_t lva = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->local_buf[0] + (cfg->jettys + i) * ctx->buf_size :
+        (uint64_t)ctx->local_buf[i] + ctx->buf_size;    // Second half for local memory
     urma_sge_t *local_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR + 1];
 
     // Step increased value
@@ -894,8 +906,12 @@ static void init_atomic_jfs_wr(urma_jfs_wr_t *wr, perftest_context_t *ctx, perft
     uint32_t remainder = (ctx->page_size / PERFTEST_BUF_NUM >= align_size) ?
         (j % ((ctx->page_size / PERFTEST_BUF_NUM) / align_size)) : 0;
 
-    uint8_t *lva = (uint8_t *)ctx->local_buf[i] + ctx->buf_size + remainder * align_size;
-    uint8_t *rva = (uint8_t *)ctx->remote_seg[i]->ubva.va + remainder * align_size;
+    uint8_t *lva = (cfg->seg_pre_jetty == false) ?
+        (uint8_t *)ctx->local_buf[0] + (cfg->jettys + i) * ctx->buf_size + remainder * align_size :
+        (uint8_t *)ctx->local_buf[i] + ctx->buf_size + remainder * align_size;
+    uint8_t *rva = (cfg->seg_pre_jetty == false) ?
+        (uint8_t *)ctx->remote_seg[i]->ubva.va + i * ctx->buf_size + remainder * align_size :
+        (uint8_t *)ctx->remote_seg[i]->ubva.va + remainder * align_size;
 
     urma_sge_t *local_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR + 1];
     urma_sge_t *remote_sge = &run_ctx->jfs_sge[(i * cfg->jfs_post_list + j) * PERFTEST_SGE_NUM_PRE_WR];
@@ -1070,7 +1086,9 @@ static void init_jfr_wr(urma_jfr_wr_t *wr, perftest_context_t *ctx, perftest_con
     run_test_ctx_t *run_ctx = &ctx->run_ctx;
 
     urma_sge_t *local_sge = &run_ctx->jfr_sge[(i * cfg->jfr_post_list + j) * PERFTEST_SGE_NUM_PRE_WR + 1];
-    local_sge->addr = (uint64_t)ctx->local_buf[i];
+    local_sge->addr = (cfg->seg_pre_jetty == false) ?
+        (uint64_t)ctx->local_buf[0] + (cfg->jettys + i) * ctx->buf_size :
+        (uint64_t)ctx->local_buf[i];
     local_sge->len = cfg->size;
     local_sge->tseg = ctx->local_tseg[i];
 
@@ -1637,7 +1655,7 @@ static int run_once_jetty_send_lat(perftest_context_t *ctx, perftest_config_t *c
                     return -1;
                 }
                 if (cr.status != URMA_CR_SUCCESS) {
-                    (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu", (int)cr.status, scnt, rcnt);
+                    (void)fprintf(stderr, "Failed CR status %d, scnt: %lu, rcnt: %lu.\n", (int)cr.status, scnt, rcnt);
                     return -1;
                 }
 
@@ -2023,7 +2041,7 @@ static int run_once_bw(perftest_context_t *ctx, perftest_config_t *cfg)
                 for (int i = 0; i < cqe_cnt; i++) {
                     cr_id = (int)cr[i].user_ctx; // todo jfs_id
                     if (cr[i].status != URMA_CR_SUCCESS) {
-                        (void)fprintf(stderr, "Failed CR status %d, tot_scnt: %lu, tot_ccnt: %lu",
+                        (void)fprintf(stderr, "Failed CR status %d, tot_scnt: %lu, tot_ccnt: %lu.\n",
                             (int)cr[i].status, tot_scnt, tot_ccnt);
                         goto free_cr;
                     }
@@ -2182,7 +2200,7 @@ static int run_once_bw_recv(perftest_context_t *ctx, perftest_config_t *cfg)
                         goto cleaning;
                     }
                     if (cr[i].status != URMA_CR_SUCCESS) {
-                        (void)fprintf(stderr, "Failed CR status %d, rcnt: %lu\n", (int)cr[i].status, rcnt);
+                        (void)fprintf(stderr, "Failed CR status %d, rcnt: %lu.\n", (int)cr[i].status, rcnt);
                         ret = -1;
                         goto cleaning;
                     }
@@ -2451,7 +2469,7 @@ static int run_once_bi_bw(perftest_context_t *ctx, perftest_config_t *cfg)
                     goto cleaning;
                 }
                 if (cr_recv[i].status != URMA_CR_SUCCESS) {
-                    (void)fprintf(stderr, "Failed CR status: %d, tot_scnt: %lu, tot_ccnt: %lu, tot_rcnt: %lu\n.",
+                    (void)fprintf(stderr, "Failed CR status: %d, tot_scnt: %lu, tot_ccnt: %lu, tot_rcnt: %lu.\n",
                         (int)cr_recv[i].status, tot_scnt, tot_ccnt, tot_rcnt);
                     ret = -1;
                     goto cleaning;