diff --git a/elements/ip/ip_route_core.cc b/elements/ip/ip_route_core.cc
index 26374df..1d41446 100644
--- a/elements/ip/ip_route_core.cc
+++ b/elements/ip/ip_route_core.cc
@@ -47,7 +47,7 @@ int nba::ipv4route::load_rib_from_file(
         uint32_t addr = ntohl(inet_addr(str_addr));
         uint16_t len = atoi(str_len);
 
-        add_route(tables, addr, len, rand() % 65536);
+        add_route(tables, addr, len, rand() % 65532 + 1);
     }
     fclose(fp);
     return 0;
diff --git a/src/lib/offloadtask.cc b/src/lib/offloadtask.cc
index 99d2717..6b3ad47 100644
--- a/src/lib/offloadtask.cc
+++ b/src/lib/offloadtask.cc
@@ -131,6 +131,7 @@ void OffloadTask::prepare_read_buffer()
                 for (PacketBatch *batch : batches) {
                     struct datablock_tracker *t = &batch->datablock_states[dbid];
                     tie(t->in_size, t->in_count) = db->calc_read_buffer_size(batch);
+                    // Now aligned_item_sizes has valid values.
                     if (t->in_size > 0 && t->in_count > 0) {
                         cctx->alloc_input_buffer(io_base, t->in_size,
                                                  t->host_in_ptr, t->dev_in_ptr);
diff --git a/tests/test_ipv4route.cc b/tests/test_ipv4route.cc
index a00c028..a01114e 100644
--- a/tests/test_ipv4route.cc
+++ b/tests/test_ipv4route.cc
@@ -97,10 +97,6 @@ TEST_P(IPLookupCUDAMatchTest, SinglePacket) {
                              ntohl(inet_addr(ip1)), &cpu_results[0]);
     ipv4route::direct_lookup(tbl24_h, tbllong_h,
                              ntohl(inet_addr(ip2)), &cpu_results[1]);
-    printf("converted ipaddr 1: 0x%08x\n", ntohl(inet_addr(ip1)));
-    printf("result 1: 0x%04x\n", cpu_results[0]);
-    printf("converted ipaddr 2: 0x%08x\n", ntohl(inet_addr(ip2)));
-    printf("result 2: 0x%04x\n", cpu_results[1]);
     EXPECT_NE(0, cpu_results[0]);
     EXPECT_NE(0, cpu_results[1]);
 
@@ -193,7 +189,7 @@ TEST_P(IPLookupCUDAMatchTest, SinglePacket) {
         &checkbits_d,
         &tbl24_d, &tbllong_d
     };
-    ASSERT_EQ(cudaSuccess, cudaLaunchKernel(k, dim3(num_batches), dim3(256),
+    ASSERT_EQ(cudaSuccess, cudaLaunchKernel(k, dim3(1), dim3(256),
               raw_args, 1024, 0));
     ASSERT_EQ(cudaSuccess, cudaDeviceSynchronize());
     ASSERT_EQ(cudaSuccess, cudaMemcpy(output_buffer, output_buffer_d,
@@ -209,6 +205,13 @@ TEST_P(IPLookupCUDAMatchTest, SinglePacket) {
     free(datablocks[1]);
     free(input_buffer);
     free(output_buffer);
+    ASSERT_EQ(cudaSuccess, cudaFree(input_buffer_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(output_buffer_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(batch_ids_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(item_ids_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(db_ipv4_dest_addrs_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(db_ipv4_lookup_results_d));
+    ASSERT_EQ(cudaSuccess, cudaFree(datablocks_d));
 }
 
 TEST_P(IPLookupCUDAMatchTest, Batch) {