diff --git a/qdp_project/src/Benchmark.cpp b/qdp_project/src/Benchmark.cpp
index 236aaf1..4600773 100644
--- a/qdp_project/src/Benchmark.cpp
+++ b/qdp_project/src/Benchmark.cpp
@@ -39,9 +39,12 @@ uint64_t* DATA_DST_;
 // if more j than b -> subsplit b like it is now
 
 void scan_b(size_t gid, size_t tid) {
-    constexpr size_t SUBCHUNK_COUNT = TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
+    constexpr size_t RUN_INCREMENT = TC_SCANB / TC_AGGRJ;
+    constexpr size_t THREAD_RATIO = TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
+    constexpr size_t SUBCHUNK_COUNT = THREAD_RATIO > 0 ? THREAD_RATIO : 1;
     constexpr size_t SUBCHUNK_SIZE_B = CHUNK_SIZE_B / SUBCHUNK_COUNT;
     constexpr size_t SUBCHUNK_SIZE_ELEMENTS = CHUNK_SIZE_ELEMENTS / SUBCHUNK_COUNT;
+    constexpr size_t TC_ACTUAL = TC_AGGRJ >= TC_SCANB ? TC_SCANB : TC_AGGRJ;
 
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)].clear();
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)].resize(1);
@@ -51,9 +54,9 @@ void scan_b(size_t gid, size_t tid) {
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)][0][TIME_STAMP_BEGIN] = std::chrono::steady_clock::now();
 
     if constexpr (PERFORM_CACHING) {
-        for (size_t i = 0; i < RUN_COUNT; i++) {
+        for (size_t i = 0; i < RUN_COUNT; i += RUN_INCREMENT) {
             const size_t chunk_index = get_chunk_index(gid, i);
-            uint64_t* chunk_ptr = get_chunk<TC_SCANB>(DATA_B_, chunk_index, tid);
+            uint64_t* chunk_ptr = get_chunk<TC_ACTUAL>(DATA_B_, chunk_index, tid);
 
             for (size_t j = 0; j < SUBCHUNK_COUNT; j++) {
                 uint64_t* sub_chunk_ptr = &chunk_ptr[j * SUBCHUNK_SIZE_ELEMENTS];
diff --git a/qdp_project/src/Configuration.hpp b/qdp_project/src/Configuration.hpp
index 66e87f2..df4de1d 100644
--- a/qdp_project/src/Configuration.hpp
+++ b/qdp_project/src/Configuration.hpp
@@ -5,10 +5,10 @@ constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
 constexpr uint32_t ITERATION_COUNT = 5;
 
 #ifdef MODE_SIMPLE_PREFETCH
-constexpr uint32_t GROUP_COUNT = 16;
+constexpr uint32_t GROUP_COUNT = 8;
 constexpr size_t CHUNK_SIZE_B = 64_MiB;
 constexpr uint32_t TC_SCANA = 2;
-constexpr uint32_t TC_SCANB = 1;
+constexpr uint32_t TC_SCANB = 2;
 constexpr uint32_t TC_AGGRJ = 1;
 constexpr bool PERFORM_CACHING = true;
 constexpr bool STORE_B_IN_HBM = false;