use 4gib as size and again 1 aggrj thread for qdp bench

qdp_project/src/Configuration.hpp

@@ -1,6 +1,6 @@
 #pragma once
 
-constexpr size_t WL_SIZE_B = 1_GiB;
+constexpr size_t WL_SIZE_B = 4_GiB;
 constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
 constexpr uint32_t ITERATION_COUNT = 5;
 
@@ -9,7 +9,7 @@ constexpr uint32_t GROUP_COUNT = 8;
 constexpr size_t CHUNK_SIZE_B = 64_MiB;
 constexpr uint32_t TC_SCANA = 1;
 constexpr uint32_t TC_SCANB = 1;
-constexpr uint32_t TC_AGGRJ = 2;
+constexpr uint32_t TC_AGGRJ = 1;
 constexpr bool PERFORM_CACHING = true;
 constexpr bool STORE_B_IN_HBM = false;
 constexpr char MODE_STRING[] = "simple-prefetch";