CarpetX - Revision history

Scupp: /* SyncTestX thorn */

2021-12-08T16:10:27Z

‎SyncTestX thorn

Scupp: /* SyncTestX thorn */

2021-12-08T15:58:28Z

‎SyncTestX thorn

Scupp: /* SyncTestX thorn */

2021-12-08T15:55:42Z

‎SyncTestX thorn

Scupp: /* SyncTestX thorn */

2021-12-08T15:54:31Z

‎SyncTestX thorn

Scupp: /* SyncTestX thorn */

2021-12-08T15:50:15Z

‎SyncTestX thorn

Scupp at 15:42, 8 December 2021

2021-12-08T15:42:59Z

Scupp at 23:33, 20 October 2021

2021-10-20T23:33:49Z

Scupp: /* Multipole/qc0 */

2021-10-20T05:15:28Z

‎Multipole/qc0

Scupp at 05:07, 20 October 2021

2021-10-20T05:07:58Z

Scupp at 05:05, 20 October 2021

2021-10-20T05:05:41Z

@@ Line 44: / Line 44: @@
 == SyncTestX thorn ==
 * Wrote SyncTestX thorn for comparing the communication overhead for Carpet and CarpetX
 * Contains parameter files (and thornlists) for a weak scaling test using the thorn
 * Ran weak scaling test on LSU's Deep Bayou. The runs were scaled from 1 to 4 nodes (48 to 192 processors), and the number of threads per process was varied from 1 to 8. Except for 8 threads, CarpetX outperformed Carpet in the tests. The data is shown in the included plots. Also shown is a plot of the CarpetX time divided by the Carpet time, clearly showing that Carpet performed better only for the runs with 8 threads (with 1,2, and 4 nodes) and 4 threads (with 1 node).
-[[File:CarpetX SyncComparison1.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads.]]
+[[File:CarpetX SyncComparison1.png|thumb|Comparison of the runtimes for the weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads.]]
-[[File:CarpetX SyncComparison2.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 4 thread and 8 threads.]]
+[[File:CarpetX SyncComparison2.png|thumb|Comparison of the runtimes for the weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 4 thread and 8 threads.]]
-[[File:SyncComparison3.png|thumb|ff]]
+[[File:SyncComparison3.png|thumb|Plot of the runtime ratio (CarpetX/Carpet) for the weak scaling test with SyncTestX. Values above 1 mean that Carpet took less time to run than CarpetX. Carpet outperformed CarpetX only in the runs with 48 processors with 4 and 8 threads, 96 processors with 8 threads, and 192 processors with 8 threads.]]
 == Open issues/bugs ==

@@ Line 53: / Line 53: @@
 [[File:SyncComparison3.png|thumb|ff]]
-{{multiple image
+[[File:CarpetX SyncComparison1.png|thumb|File:CarpetX SyncComparison2.png|thumb]]
 == Open issues/bugs ==

@@ Line 49: / Line 49: @@
 [[File:CarpetX SyncComparison1.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads.]]
-[[File:CarpetX SyncComparison2.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads. CarpetX outperforms Carpet in all of these runs.]]
+[[File:CarpetX SyncComparison2.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 4 thread and 8 threads.]]
 [[File:SyncComparison3.png|thumb|ff]]
-<gallery>
+{{multiple image
-CarpetX SyncComparison1.png|Caption1
+ | width = 60
-CarpetX SyncComparison2.png|Caption2
+ | image1 = CarpetX SyncComparison1.png
-</gallery>
+ | image2 = CarpetX SyncComparison2.png
 == Open issues/bugs ==

←Older revision		Revision as of 15:54, 8 December 2021
Line 52:		Line 52:

	[[File:SyncComparison3.png\|thumb\|ff]]		[[File:SyncComparison3.png\|thumb\|ff]]
		+
		+	<gallery>
		+	CarpetX SyncComparison1.png\|Caption1
		+	CarpetX SyncComparison2.png\|Caption2
		+	</gallery>

	== Open issues/bugs ==		== Open issues/bugs ==

@@ Line 47: / Line 47: @@
 * Ran weak scaling test on LSU's Deep Bayou. The runs were scaled from 1 to 4 nodes (48 to 192 processors), and the number of threads per process was varied from 1 to 8. Except for 8 threads, CarpetX outperformed Carpet in the tests. The data is shown in the included plots. Also shown is a plot of the CarpetX time divided by the Carpet time, clearly showing that Carpet performed better only for the runs with 8 threads (with 1,2, and 4 nodes) and 4 threads (with 1 node).
-[[File:Psi4.png|thumb|Plot of the L2,m2 mode output for Psi4 at r=10. This data is from the 200M simulation ran on Melete.]]
+[[File:CarpetX SyncComparison1.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads.]]
-[[File:CarpetX SyncComparison1.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads. CarpetX outperforms Carpet in all of these runs.]]
+[[File:CarpetX SyncComparison2.png|thumb|Comparison of the runtimes for a weak scaling test for Carpet and CarpetX using the SyncTestX thorn. The test scales from 48 to 192 processors, and this plot shows the results for runs with 1 thread and 2 threads. CarpetX outperforms Carpet in all of these runs.]]
 == Open issues/bugs ==

@@ Line 41: / Line 41: @@
 | 40 || 358.8 || 527.4
 |}
 == Open issues/bugs ==

@@ Line 25: / Line 25: @@
 [[File:Psi4.png|thumb|Plot of the L2,m2 mode output for Psi4 at r=10. This data is from the 200M simulation ran on Melete.]]
-* Recent testing of just synchronization speeds (branched off of a more recent master) showed similar speeds, so I ran my previous comparison test with Minkowski initial data. The timing data from those runs is shown below. The runs are much closer than before, so a change in master has fixed much of the speed issue. CarpetX still takes 30-50% more time than Carpet, but it is far better than before.
+* Recent testing of just synchronization speeds (branched off of a more recent master) showed similar speeds, so I ran my previous comparison test with Minkowski initial data using the newer master. The timing data from those runs is shown below. The runs are much closer than before, so a change in master has fixed much of the speed issue. CarpetX still takes 30-50% more time than Carpet, but it is far better than before.
 {| class="wikitable"