RTWG-discussion-points - Revision history

Peter Huewe: spellcheck

2011-04-11T21:01:39Z

spellcheck

2007-09-04T07:40:18Z

‎shared Test Bed ideas

2007-09-04T07:36:02Z

2007-09-04T06:59:37Z

‎table of test attributes

2007-09-04T06:50:42Z

2007-09-04T05:57:59Z

‎cyclictest

2007-09-04T05:56:02Z

‎cyclictest

2007-09-04T05:44:08Z

2007-09-04T05:38:54Z

2007-09-04T05:16:39Z

@@ Line 96: / Line 96: @@
 page 26 - send performance of nonRT kernel is used as baseline for normalization.
 Measurement is of network performance (with several different configurations).
-Result of good performance for non-RT kernel, but with interupt threading is
+Result of good performance for non-RT kernel, but with interrupt threading is
 not expected.
@@ Line 187: / Line 187: @@
 |serial port
 |lpar interrupt (hypervisor)
-|timer system (reprogrammed inter intterrupt)
+|timer system (reprogrammed inter interrupt)
 |-
 |# machines

← Older revision		Revision as of 07:40, 4 September 2007
Line 231:		Line 231:
	=== shared Test Bed ideas ===		=== shared Test Bed ideas ===
	See Darren Hart's tests at : http://www.dvhart.com/~dvhart/ols2007		See Darren Hart's tests at : http://www.dvhart.com/~dvhart/ols2007
		+
		+	His paper is at: https://ols2006.108.redhat.com/2007/Reprints/rostedt-Reprint.pdf

← Older revision		Revision as of 07:36, 4 September 2007
Line 227:		Line 227:
	\|no, but needs HRT support for arch.		\|no, but needs HRT support for arch.
	\|}		\|}
		+
		+
		+	=== shared Test Bed ideas ===
		+	See Darren Hart's tests at : http://www.dvhart.com/~dvhart/ols2007

@@ Line 157: / Line 157: @@
 == table of test attributes ==
-{|style="border:1px;"
+{|border="1" cellspacing="0" cellpadding="5"
-|-
+|- style="background:#CCCCCC"
 |attribute / test->
 |rf-etri
@@ Line 182: / Line 182: @@
 |excel charts
 |none
 |}

← Older revision		Revision as of 06:50, 4 September 2007
Line 155:		Line 155:

	don't know how to interpret results (don't know what constitutes good values)		don't know how to interpret results (don't know what constitutes good values)
		+
		+	== table of test attributes ==
		+	{\|style="border:1px;"
		+	\|-
		+	\|attribute / test->
		+	\|rf-etri
		+	\|igel
		+	\|celleb
		+	\|cyclictest
		+	\|-
		+	\|platform specific?
		+	\|yes - uses TSC, requires /dev/rtc
		+	\|yes - uses SH timestamp
		+	\|yes - uses lpar interrupt, uses cell timebase regs.
		+	\|no, but requires HRT support. Uses linux time system for timestamps
		+	\|-
		+	\|data format/handling
		+	\|text->csv->excel
		+	\|csv text
		+	\|text->csv->excel
		+	\|none
		+	\|-
		+	\|presentation
		+	\|excel charts
		+	\|gnuplot
		+	\|excel charts
		+	\|none
		+	\|}

@@ Line 139: / Line 139: @@
 benefits:
 * seems to be what other people are using.
 problems:
@@ Line 145: / Line 147: @@
 * requires newer structure or API for timer info (not on MIPS)
 * it doesn't have kernel instrumentation point
 Notes:

@@ Line 144: / Line 144: @@
 * requires newer version of glibc and compiler
 * requires newer structure or API for timer info (not on MIPS)
 Notes:
@@ Line 149: / Line 150: @@
 * it works on ppc64
 * it didn't work on MIPS for Tim

← Older revision		Revision as of 05:44, 4 September 2007
Line 133:		Line 133:
	* is better than realfeel, because it measures time from ISR start		* is better than realfeel, because it measures time from ISR start
	* uses read-unblock, which has good code path through kernel.		* uses read-unblock, which has good code path through kernel.
		+
		+	=== cyclictest ===
		+	User timestamps from timer system to measure performance of timer system.
		+
		+	benefits:
		+	* seems to be what other people are using.
		+
		+	problems:
		+	* requires HRT support
		+	* requires newer version of glibc and compiler
		+	* requires newer structure or API for timer info (not on MIPS)
		+
		+	Notes:
		+	* it works on SH, on 2.6.22 or later.
		+	* it works on ppc64
		+	* it didn't work on MIPS for Tim

← Older revision		Revision as of 05:38, 4 September 2007
Line 113:		Line 113:
	* wanted to see if RT-preempt affected throughput performance		* wanted to see if RT-preempt affected throughput performance
	** result: it does affect performance some, but it is not a big effect.		** result: it does affect performance some, but it is not a big effect.
		+
		+	=== etri testing ===
		+	* wanted to compare RT-preempt with regular kernel.
		+	** worst case latency was good with RT-preempt patch
		+	* Wanted to use modified realfeel code (realfeel-etri)
		+	** it worked.
		+	* wanted to know relationship between scheduling latency and throughput performance
		+	** result: there were possible throughput problems with the RT-preempt kernel
		+
		+	* haven't used realfeel-etri on ARM, yet.
		+
		+	=== realfeel-etri ===
		+	problems:
		+	* uses TSC (but abstracts the get_ticks() interface
		+	** have to use something else on ARM
		+	* needs /dev/RTC (may not be available on other embedded platforms)
		+
		+	benefits:
		+	* is better than realfeel, because it measures time from ISR start
		+	* uses read-unblock, which has good code path through kernel.

← Older revision		Revision as of 05:16, 4 September 2007
Line 107:		Line 107:
	of longer duration interrupt-off times, the network card may buffer packets, and the		of longer duration interrupt-off times, the network card may buffer packets, and the
	network stack may be able to process more data at a time.		network stack may be able to process more data at a time.
		+
		+	Tried to test:
		+	* did RT-preempt patch for PPC64 work?
		+	** result: yes, to some extent
		+	* wanted to see if RT-preempt affected throughput performance
		+	** result: it does affect performance some, but it is not a big effect.