diff --git a/kernel/ap.c b/kernel/ap.c
index 1105672a..8fa7dda5 100644
--- a/kernel/ap.c
+++ b/kernel/ap.c
@@ -68,6 +68,11 @@ static void ap_wait(void)
 		init_host_ikc2mckernel();
 		init_host_ikc2linux(ikc_cpu);
 		mcs_lock_unlock_noirq(&ap_syscall_semaphore, &mcs_node);
+
+		{
+			extern void hfi1_kmalloc_cache_prealloc(void);
+			hfi1_kmalloc_cache_prealloc();
+		}
 	}
 	
 	/* one of them listens */
diff --git a/kernel/include/cls.h b/kernel/include/cls.h
index a16523af..a0ffa413 100644
--- a/kernel/include/cls.h
+++ b/kernel/include/cls.h
@@ -106,7 +106,12 @@ struct cpu_local_var {
 	struct list_head smp_func_req_list;
 
 	struct process_vm *on_fork_vm;
+
+	/* HFI1 related per-core kmalloc caches */
 	struct kmalloc_cache_header txreq_cache;
+	struct kmalloc_cache_header tids_cache;
+	struct kmalloc_cache_header tidlist_cache;
+	struct kmalloc_cache_header tid_node_cache;
 } __attribute__((aligned(64)));
 
 
diff --git a/kernel/init.c b/kernel/init.c
index a5d4890f..7d68060e 100644
--- a/kernel/init.c
+++ b/kernel/init.c
@@ -360,6 +360,11 @@ static void post_init(void)
 		}
 		init_host_ikc2mckernel();
 		init_host_ikc2linux(ikc_cpu);
+
+		{
+			extern void hfi1_kmalloc_cache_prealloc(void);
+			hfi1_kmalloc_cache_prealloc();
+		}
 	}
 
 	arch_setup_vdso();
diff --git a/kernel/syscall.c b/kernel/syscall.c
index cad6edc0..854e69b9 100644
--- a/kernel/syscall.c
+++ b/kernel/syscall.c
@@ -484,14 +484,12 @@ long do_syscall(struct syscall_request *req, int cpu, int pid)
 		if ((cpu_local_var(current)->proc->mcexec_flags & MCEXEC_HFI1) &&
 				res.private_data &&
 				!strncmp((const char *)req->args[0], "/dev/hfi", 8)) {
-			extern void hfi1_txreq_prealloc(void);
 
 			thread->proc->fd_priv_table[rc] = res.private_data;
 			dkprintf("%s: PID: %d, open fd: %d, filename: "
 					"%s, private_data: 0x%lx\n",
 					__FUNCTION__, thread->proc->pid,
 					rc, req->args[0], res.private_data);
-			hfi1_txreq_prealloc();
 		}
 	}
 
diff --git a/kernel/user_exp_rcv.c b/kernel/user_exp_rcv.c
index 39325908..b0582327 100644
--- a/kernel/user_exp_rcv.c
+++ b/kernel/user_exp_rcv.c
@@ -67,8 +67,6 @@ static int set_rcvarray_entry(struct hfi1_filedata *, uintptr_t,
 static int unprogram_rcvarray(struct hfi1_filedata *, u32, struct tid_group **);
 static void clear_tid_node(struct hfi1_filedata *, struct tid_rb_node *);
 
-struct kmalloc_cache_header tidlist_cache = { NULL };
-
 /*
  * RcvArray entry allocation for Expected Receives is done by the
  * following algorithm:
@@ -93,9 +91,10 @@ int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
 		return -EINVAL;
 	}
 
-	tidlist = kmalloc_cache_alloc(&tidlist_cache,
-			//sizeof(*tidlist) * uctxt->expected_count);
-			sizeof(*tidlist) * 1024);
+	/* TODO: sizeof(*tidlist) * uctxt->expected_count); */
+	tidlist = kmalloc_cache_alloc(&cpu_local_var(tidlist_cache),
+			sizeof(*tidlist) * 2048);
+
 	if (!tidlist)
 		return -ENOMEM;
 
@@ -328,7 +327,6 @@ static int program_rcvarray(struct hfi1_filedata *fd, uintptr_t phys,
 	return count;
 }
 
-struct kmalloc_cache_header tid_node_cache = { NULL };
 
 static int set_rcvarray_entry(struct hfi1_filedata *fd, uintptr_t phys,
 			      u32 rcventry, struct tid_group *grp,
@@ -342,7 +340,8 @@ static int set_rcvarray_entry(struct hfi1_filedata *fd, uintptr_t phys,
 	 * Allocate the node first so we can handle a potential
 	 * failure before we've programmed anything.
 	 */
-	node = kmalloc_cache_alloc(&tid_node_cache, sizeof(*node));
+	node = kmalloc_cache_alloc(&cpu_local_var(tid_node_cache),
+			sizeof(*node));
 	if (!node)
 		return -ENOMEM;
 
diff --git a/kernel/user_sdma.c b/kernel/user_sdma.c
index 3dbadb9f..4eba3ab8 100644
--- a/kernel/user_sdma.c
+++ b/kernel/user_sdma.c
@@ -873,8 +873,6 @@ int hfi1_unmap_device_addresses(struct process *proc)
 	return ret;
 }
 
-struct kmalloc_cache_header tids_cache = {NULL};
-
 #undef PROFILE_ENABLE
 
 #ifdef __HFI1_ORIG__
@@ -1143,7 +1141,8 @@ int hfi1_user_sdma_process_request(void *private_data, struct iovec *iovec,
 			ret = -EINVAL;
 			goto free_req;
 		}
-		req->tids = kmalloc_cache_alloc(&tids_cache,
+		req->tids = kmalloc_cache_alloc(
+				&cpu_local_var(tids_cache),
 				sizeof(*req->tids) * MAX_TID_PAIR_ENTRIES);
 		if (!req->tids) {
 			ret = -ENOMEM;
@@ -1336,10 +1335,22 @@ static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
 }
 
-void hfi1_txreq_prealloc(void)
+void hfi1_kmalloc_cache_prealloc(void)
 {
+	/*
+	 * TODO: nr_elems have been determined based on profiling
+	 * HACC and UMT2013, would be interesting to do some clever
+	 * dynamic releasing/expanding.
+	 */
 	kmalloc_cache_prealloc(&cpu_local_var(txreq_cache),
-			sizeof(struct user_sdma_txreq));
+			sizeof(struct user_sdma_txreq), 2048);
+	kmalloc_cache_prealloc(&cpu_local_var(tids_cache),
+			sizeof(*(((struct user_sdma_request *)0)->tids)) *
+				MAX_TID_PAIR_ENTRIES, 256);
+	kmalloc_cache_prealloc(&cpu_local_var(tidlist_cache),
+			sizeof(u32) * 2048, 128);
+	kmalloc_cache_prealloc(&cpu_local_var(tid_node_cache),
+			sizeof(struct tid_rb_node), 512);
 }
 
 static int user_sdma_send_pkts(struct user_sdma_request *req,