diff --git a/cuda/cuda_util.c b/cuda/cuda_util.c
index c0e5c6e5..3fe61cc0 100644
--- a/cuda/cuda_util.c
+++ b/cuda/cuda_util.c
@@ -37,11 +37,19 @@ static int hasUVA=-1;
 static struct cudaDeviceProp *prop=NULL;
 static spgpuHandle_t psb_cuda_handle = NULL;
 static cublasHandle_t psb_cublas_handle = NULL;
-
+#if defined(TRACK_CUDA_MALLOC)
+static long long total_cuda_mem = 0;
+#endif
 
 int allocRemoteBuffer(void** buffer, int count)
 {
   cudaError_t err = cudaMalloc(buffer, count);
+#if defined(TRACK_CUDA_MALLOC)
+  total_cuda_mem += count;
+  fprintf(stderr,"Tracking CUDA allocRemoteBuffer for %ld bytes total  %ld  address %p\n",
+	  count, total_cuda_mem, *buffer);
+#endif
+    
   if (err == cudaSuccess)
     {
       return SPGPU_SUCCESS;
diff --git a/cuda/fcusparse_fct.h b/cuda/fcusparse_fct.h
index 06facdc0..689bdc93 100644
--- a/cuda/fcusparse_fct.h
+++ b/cuda/fcusparse_fct.h
@@ -178,7 +178,8 @@ int T_spmvCSRGDevice(T_Cmat *Matrix, TYPE alpha, void *deviceX,
       CHECK_CUDA(cudaFree(cMat->mvbuffer));
       cMat->mvbuffer = NULL;
     }
-    CHECK_CUDA(cudaMalloc((void **) &(cMat->mvbuffer), bfsz));
+    //CHECK_CUDA(cudaMalloc((void **) &(cMat->mvbuffer), bfsz));
+    allocRemoteBuffer((void **) &(cMat->mvbuffer), bfsz);
     cMat->mvbsize = bfsz;
   }
   CHECK_CUSPARSE(cusparseCsrmvEx(*my_handle,
@@ -215,7 +216,9 @@ int T_spmvCSRGDevice(T_Cmat *Matrix, TYPE alpha, void *deviceX,
       CHECK_CUDA(cudaFree(cMat->mvbuffer));
       cMat->mvbuffer = NULL;
     }
-    CHECK_CUDA(cudaMalloc((void **) &(cMat->mvbuffer), bfsz));
+    //CHECK_CUDA(cudaMalloc((void **) &(cMat->mvbuffer), bfsz));
+    allocRemoteBuffer((void **) &(cMat->mvbuffer), bfsz);
+    
     cMat->mvbsize = bfsz;
   }
   CHECK_CUSPARSE(cusparseSpMV(*my_handle,CUSPARSE_OPERATION_NON_TRANSPOSE,
@@ -287,7 +290,9 @@ int T_spsvCSRGDevice(T_Cmat *Matrix, TYPE alpha, void *deviceX,
 	CHECK_CUDA(cudaFree(cMat->svbuffer));
 	cMat->svbuffer = NULL;
       }
-      CHECK_CUDA(cudaMalloc((void **) &(cMat->svbuffer), bfsz));
+      //CHECK_CUDA(cudaMalloc((void **) &(cMat->svbuffer), bfsz));
+      allocRemoteBuffer((void **) &(cMat->svbuffer), bfsz);
+    
       cMat->svbsize=bfsz;
       CHECK_CUSPARSE(cusparseSpSV_analysis(*my_handle,
 					   CUSPARSE_OPERATION_NON_TRANSPOSE,
@@ -382,7 +387,9 @@ int T_CSRGDeviceAlloc(T_Cmat *Matrix,int nr, int nc, int nz)
   /*   cMat->svbuffer = NULL; */
   /* } */
   if (bfsz > 0) {
-    CHECK_CUDA(cudaMalloc((void **) &(cMat->svbuffer), bfsz));
+    //CHECK_CUDA(cudaMalloc((void **) &(cMat->svbuffer), bfsz));
+    allocRemoteBuffer((void **) &(cMat->svbuffer), bfsz);
+
   } else {
     cMat->svbuffer=NULL;
   }