summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorRaymond Lu <songyulu@hdfgroup.org>2009-02-11 18:50:21 (GMT)
committerRaymond Lu <songyulu@hdfgroup.org>2009-02-11 18:50:21 (GMT)
commit7768b49c16491ee9936ca40e375410121b44e15a (patch)
treeef7fa141bb66db80e517cb7b2c67df92a79b84df
parent725b4d7ddb461c984c8afef2f7ea5d962f5def53 (diff)
downloadhdf5-7768b49c16491ee9936ca40e375410121b44e15a.zip
hdf5-7768b49c16491ee9936ca40e375410121b44e15a.tar.gz
hdf5-7768b49c16491ee9936ca40e375410121b44e15a.tar.bz2
[svn-r16466] Performance Improvement(bug #1450). When a chunk is bigger than the cache size and isn't allocated
on disk, the library still loaded it in the cache, which is redundant. I changed it to bypass the cache and added a test in dsets.c. Tested on jam and smirom.
-rw-r--r--test/dsets.c135
1 files changed, 135 insertions, 0 deletions
diff --git a/test/dsets.c b/test/dsets.c
index e3bac54..575449a 100644
--- a/test/dsets.c
+++ b/test/dsets.c
@@ -44,6 +44,7 @@ const char *FILENAME[] = {
"random_chunks",
"huge_chunks",
"chunk_cache",
+ "big_chunk",
NULL
};
#define FILENAME_BUF_SIZE 1024
@@ -163,6 +164,12 @@ const char *FILENAME[] = {
#define TOO_HUGE_CHUNK_DIM2_1 ((hsize_t)1024)
#define TOO_HUGE_CHUNK_DIM2_2 ((hsize_t)1024)
+/* Parameters for testing bypassing chunk cache */
+#define BYPASS_DATASET "Dset"
+#define BYPASS_DIM 1000
+#define BYPASS_CHUNK_DIM 500
+#define BYPASS_FILL_VALUE 7
+
/* Shared global arrays */
#define DSET_DIM1 100
#define DSET_DIM2 200
@@ -6683,6 +6690,133 @@ error:
/*-------------------------------------------------------------------------
+ * Function: test_big_chunks_bypass_cache
+ *
+ * Purpose: When the chunk size is bigger than the cache size and the
+ * chunk isn't on disk, this test verifies that the library
+ * bypasses the cache.
+ *
+ * Return: Success: 0
+ * Failure: -1
+ *
+ * Programmer: Raymond Lu
+ * 11 Feb 2009
+ *
+ *-------------------------------------------------------------------------
+ */
+static herr_t
+test_big_chunks_bypass_cache(hid_t fapl)
+{
+ char filename[FILENAME_BUF_SIZE];
+ hid_t fid = -1; /* File ID */
+ hid_t dcpl = -1; /* Dataset creation property list ID */
+ hid_t sid = -1; /* Dataspace ID */
+ hid_t dsid = -1; /* Dataset ID */
+ hsize_t dim, chunk_dim; /* Dataset and chunk dimensions */
+ size_t rdcc_nelmts, rdcc_nbytes;
+ int fvalue = BYPASS_FILL_VALUE;
+ hsize_t count, stride, offset, block;
+ static int wdata[BYPASS_CHUNK_DIM], rdata[BYPASS_DIM];
+ int i, j;
+ herr_t ret; /* Generic return value */
+
+ TESTING("big chunks bypassing the cache");
+
+ h5_fixname(FILENAME[9], fapl, filename, sizeof filename);
+
+ /* Define cache size to be smaller than chunk size */
+ rdcc_nelmts = BYPASS_CHUNK_DIM/5;
+ rdcc_nbytes = sizeof(int)*BYPASS_CHUNK_DIM/5;
+ if(H5Pset_cache(fapl, 0, rdcc_nelmts, rdcc_nbytes, 0) < 0) FAIL_STACK_ERROR
+
+ /* Create file */
+ if((fid = H5Fcreate(filename, H5F_ACC_TRUNC, H5P_DEFAULT, fapl)) < 0) FAIL_STACK_ERROR
+
+ /* Create 1-D dataspace */
+ dim = BYPASS_DIM;
+ if((sid = H5Screate_simple(1, &dim, NULL)) < 0) FAIL_STACK_ERROR
+
+ /* Create dataset creation property list */
+ if((dcpl = H5Pcreate(H5P_DATASET_CREATE)) < 0) FAIL_STACK_ERROR
+
+ /* Define chunk size. There will be only 2 chunks in the dataset. */
+ chunk_dim = BYPASS_CHUNK_DIM;
+ if(H5Pset_chunk(dcpl, 1, &chunk_dim) < 0) FAIL_STACK_ERROR
+
+ /* Define fill value, fill time, and chunk allocation time */
+ if(H5Pset_fill_value(dcpl, H5T_NATIVE_INT, &fvalue) < 0) FAIL_STACK_ERROR
+
+ if(H5Pset_fill_time(dcpl, H5D_FILL_TIME_IFSET) < 0) FAIL_STACK_ERROR
+
+ if(H5Pset_alloc_time(dcpl, H5D_ALLOC_TIME_INCR) < 0) FAIL_STACK_ERROR
+
+ /* Try to create dataset */
+ dsid = H5Dcreate2(fid, BYPASS_DATASET, H5T_NATIVE_INT, sid, H5P_DEFAULT, dcpl, H5P_DEFAULT);
+
+ /* Select first chunk to write the data */
+ offset = 0;
+ count = 1;
+ stride = 1;
+ block = BYPASS_CHUNK_DIM;
+ if(H5Sselect_hyperslab(sid, H5S_SELECT_SET, &offset, &stride, &count, &block) < 0)
+ FAIL_STACK_ERROR
+
+ for(i = 0; i < BYPASS_CHUNK_DIM; i++)
+ wdata[i] = i;
+
+ /* This write should bypass the cache because the chunk is bigger than the cache size
+ * and it's not allocated on disk. */
+ if(H5Dwrite(dsid, H5T_NATIVE_INT, H5S_ALL, sid, H5P_DEFAULT, wdata) < 0)
+ FAIL_STACK_ERROR
+
+ if(H5Dclose(dsid) < 0) FAIL_STACK_ERROR
+
+ /* Reopen the dataset */
+ if((dsid = H5Dopen(fid, BYPASS_DATASET, H5P_DEFAULT)) < 0) FAIL_STACK_ERROR
+
+ /* Reads both 2 chunks. Reading the second chunk should bypass the cache because the
+ * chunk is bigger than the cache size and it isn't allocated on disk. */
+ if(H5Dread(dsid, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, H5P_DEFAULT, rdata) < 0)
+ FAIL_STACK_ERROR
+
+ for(i = 0; i < BYPASS_CHUNK_DIM; i++)
+ if(rdata[i] != i) {
+ printf(" Read different values than written in the 1st chunk.\n");
+ printf(" At line %d and index %d, rdata = %d. It should be %d.\n", __LINE__,
+ i, rdata[i], i);
+ FAIL_STACK_ERROR
+ }
+
+ for(j = BYPASS_CHUNK_DIM; j < BYPASS_DIM; j++)
+ if(rdata[j] != fvalue) {
+ printf(" Read different values than written in the 2nd chunk.\n");
+ printf(" At line %d and index %d, rdata = %d. It should be %d.\n", __LINE__,
+ i, rdata[i], fvalue);
+ FAIL_STACK_ERROR
+ }
+
+ /* Close IDs */
+ if(H5Sclose(sid) < 0) FAIL_STACK_ERROR
+ if(H5Dclose(dsid) < 0) FAIL_STACK_ERROR
+ if(H5Pclose(dcpl) < 0) FAIL_STACK_ERROR
+ if(H5Fclose(fid) < 0) FAIL_STACK_ERROR
+
+ PASSED();
+ return 0;
+
+error:
+ H5E_BEGIN_TRY {
+ H5Pclose(dcpl);
+ H5Dclose(dsid);
+ H5Sclose(sid);
+ H5Fclose(fid);
+ } H5E_END_TRY;
+ return -1;
+} /* end test_huge_chunks() */
+
+
+
+/*-------------------------------------------------------------------------
* Function: main
*
* Purpose: Tests the dataset interface (H5D)
@@ -6803,6 +6937,7 @@ main(void)
#endif /* H5_NO_DEPRECATED_SYMBOLS */
nerrors += (test_huge_chunks(my_fapl) < 0 ? 1 : 0);
nerrors += (test_chunk_cache(my_fapl) < 0 ? 1 : 0);
+ nerrors += (test_big_chunks_bypass_cache(my_fapl) < 0 ? 1 : 0);
if(H5Fclose(file) < 0)
goto error;