diff --git a/mpp/shmemx_c_func.h4 b/mpp/shmemx_c_func.h4
index 948146156..fb3e3f720 100644
--- a/mpp/shmemx_c_func.h4
+++ b/mpp/shmemx_c_func.h4
@@ -105,3 +105,7 @@ SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_create(void *base, size_t size
 SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_preinit(void);
 SHMEM_FUNCTION_ATTRIBUTES int SHPRE()shmemx_heap_preinit_thread(int requested, int *provided);
 SHMEM_FUNCTION_ATTRIBUTES void SHPRE()shmemx_heap_postinit(void);
+
+/* Memory Ordering Routines */
+SHMEM_FUNCTION_ATTRIBUTES void shmem_pe_quiet(const int *target_pes, int npes);
+SHMEM_FUNCTION_ATTRIBUTES void shmem_ctx_pe_quiet(shmem_ctx_t ctx, const int *target_pes, int npes);
diff --git a/src/shmem_synchronization.h b/src/shmem_synchronization.h
index 0270d6d7b..e93a6226f 100644
--- a/src/shmem_synchronization.h
+++ b/src/shmem_synchronization.h
@@ -40,6 +40,25 @@ shmem_internal_quiet(shmem_ctx_t ctx)
     shmem_transport_syncmem();
 }
 
+static inline void
+shmem_internal_pe_quiet(shmem_ctx_t ctx, const int *target_pes, int npes)
+{
+    int ret;
+
+    if (ctx == SHMEM_CTX_INVALID)
+        return;
+
+    ret = shmem_transport_pe_quiet((shmem_transport_ctx_t *)ctx, target_pes, npes);
+    if (0 != ret) { RAISE_ERROR(ret); }
+
+    shmem_internal_membar();
+
+    /* Transport level memory flush is required to make memory 
+     * changes (i.e. subsequent coherent load operations 
+     * performed via the shmem_ptr API, the result of atomics 
+     * that targeted the local process) visible */
+    shmem_transport_syncmem();
+}
 
 static inline void
 shmem_internal_fence(shmem_ctx_t ctx)
diff --git a/src/synchronization_f.c b/src/synchronization_f.c
index 74e9c5c4d..30f48ec6b 100644
--- a/src/synchronization_f.c
+++ b/src/synchronization_f.c
@@ -37,6 +37,28 @@ FC_SHMEM_QUIET(void)
     shmem_internal_quiet(SHMEM_CTX_DEFAULT);
 }
 
+#define FC_SHMEM_PE_QUIET FC_FUNC_(shmem_pe_quiet, SHMEM_PE_QUIET)
+void SHMEM_FUNCTION_ATTRIBUTES
+FC_SHMEM_PE_QUIET(const int *target_pes, int npes);
+void
+FC_SHMEM_PE_QUIET(const int *target_pes, int npes)
+{
+    SHMEM_ERR_CHECK_INITIALIZED();
+
+    shmem_internal_pe_quiet(SHMEM_CTX_DEFAULT, target_pes, npes);
+}
+
+#define FC_SHMEM_CTX_QUIET FC_FUNC_(shmem_ctx_pe_quiet, SHMEM_CTX_PE_QUIET)
+void SHMEM_FUNCTION_ATTRIBUTES
+FC_SHMEM_CTX_PE_QUIET(shmem_ctx_t ctx, const int *target_pes, int npes);
+void
+FC_SHMEM_CTX_PE_QUIET(shmem_ctx_t ctx, const int *target_pes, int npes)
+{
+    SHMEM_ERR_CHECK_INITIALIZED();
+
+    shmem_internal_pe_quiet(ctx, target_pes, npes);
+}
+
 
 #define FC_SHMEM_FENCE FC_FUNC_(shmem_fence, SHMEM_FENCE)
 void SHMEM_FUNCTION_ATTRIBUTES
diff --git a/src/transport_none.h b/src/transport_none.h
index 9a10ea52c..8da3c6d98 100644
--- a/src/transport_none.h
+++ b/src/transport_none.h
@@ -103,6 +103,13 @@ shmem_transport_quiet(shmem_transport_ctx_t* ctx)
     return 0;
 }
 
+static inline
+int
+shmem_transport_pe_quiet(shmem_transport_ctx_t* ctx, const int *target_pes, int npes)
+{
+    return 0;
+}
+
 static inline
 int
 shmem_transport_fence(shmem_transport_ctx_t* ctx)
diff --git a/src/transport_ofi.c b/src/transport_ofi.c
index f8f04d65e..46fed08e2 100644
--- a/src/transport_ofi.c
+++ b/src/transport_ofi.c
@@ -2050,6 +2050,7 @@ int shmem_transport_ctx_create(struct shmem_internal_team_t *team, long options,
 #ifndef USE_CTX_LOCK
     shmem_internal_cntr_write(&ctxp->pending_put_cntr, 0);
     shmem_internal_cntr_write(&ctxp->pending_get_cntr, 0);
+    
 #endif
 
     ctxp->stx_idx = -1;
diff --git a/src/transport_ofi.h b/src/transport_ofi.h
index f7e2ebf32..b4909697d 100644
--- a/src/transport_ofi.h
+++ b/src/transport_ofi.h
@@ -548,6 +548,15 @@ int shmem_transport_quiet(shmem_transport_ctx_t* ctx)
     return 0;
 }
 
+static inline
+int shmem_transport_pe_quiet(shmem_transport_ctx_t* ctx, const int *target_pes, int npes)
+{
+
+    shmem_transport_quiet(ctx);
+
+    return 0;
+}
+
 
 static inline
 int shmem_transport_fence(shmem_transport_ctx_t* ctx)
@@ -627,7 +636,7 @@ void shmem_transport_put_scalar(shmem_transport_ctx_t* ctx, void *target, const
 
     SHMEM_TRANSPORT_OFI_CTX_LOCK(ctx);
     SHMEM_TRANSPORT_OFI_CNTR_INC(&ctx->pending_put_cntr);
-
+    
     do {
 
         ret = fi_inject_write(ctx->ep,
@@ -1002,7 +1011,6 @@ void shmem_transport_get_wait(shmem_transport_ctx_t* ctx)
     SHMEM_TRANSPORT_OFI_CTX_UNLOCK(ctx);
 }
 
-
 static inline
 void shmem_transport_cswap_nbi(shmem_transport_ctx_t* ctx, void *target, const
                                void *source, void *dest, const void *operand,
diff --git a/src/transport_portals4.h b/src/transport_portals4.h
index ab99becaa..bd3119d5d 100644
--- a/src/transport_portals4.h
+++ b/src/transport_portals4.h
@@ -292,6 +292,13 @@ shmem_transport_quiet(shmem_transport_ctx_t* ctx)
     return 0;
 }
 
+static inline
+int
+shmem_transport_pe_quiet(shmem_transport_ctx_t* ctx, const int *target_pes, int npes)
+{
+    shmem_transport_quiet(ctx);
+}
+
 
 static inline
 int
diff --git a/src/transport_ucx.h b/src/transport_ucx.h
index 2e12d22b2..cf081d7c0 100644
--- a/src/transport_ucx.h
+++ b/src/transport_ucx.h
@@ -210,6 +210,22 @@ shmem_transport_quiet(shmem_transport_ctx_t* ctx)
     return 0;
 }
 
+static inline
+int
+shmem_transport_pe_quiet(shmem_transport_ctx_t* ctx, const int *target_pes, int npes)
+{
+    ucs_status_t status;
+    
+    for (int i = 0; i < npes; i++) {
+        int pe = target_pes[i];
+        
+        status = ucp_ep_flush(shmem_transport_peers[pe].ep);
+        UCX_CHECK_STATUS(status);
+    }
+
+    return 0;
+}
+
 static inline
 int
 shmem_transport_fence(shmem_transport_ctx_t* ctx)