[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference wonder::turbolaser

Title:	TurboLaser Notesfile - AlphaServer 8200 and 8400 systems
Notice:	Welcome to WONDER::TURBOLASER in it's new homeshortly
Moderator:	LANDO::DROBNER

Created:	Tue Dec 20 1994
Last Modified:	Fri Jun 06 1997
Last Successful Update:	Fri Jun 06 1997
Number of topics:	1218
Total number of notes:	4645

1170.0. "I/O handled on more than 1 CPU ???" by BIS1::CALLEWAERT () Tue Apr 15 1997 11:50

    My customer/OEM has used AlphaServer 2100 (and got bad performance 
    because all I/O were handled by CPU 0) running UNIX.
    
    He wants to bid an 8xxx for a big project and wants to be sure that I/O 
    can be handled by more than one CPU. Can someone confirm ?
    
    Tkx.                                          
    
    Jean.

T.R	Title	User	Personal Name	Date	Lines
1170.1	version of UNIX and file system?	TROOA::MSCHNEIDER	[email protected]	`Tue Apr 15 1997 11:59`	4
	This sounds more like a UNIX issue .... with the 2100 were they using AdvFS filesystem on pre V4 Digital UNIX? If so AdvFS file system was not multithreaded prior to V4 and hence multiple CPUs did not improve I/O capabilities.
1170.2	Another reason ???	BIS1::CALLEWAERT		`Wed Apr 16 1997 12:05`	3
	It did not seem to be linked with Advfs. They are using UNIX 3.2G on their 2100. And they made their tests on 4100 with 3.2G - with REAL improvements.
1170.3	A 4100 is much faster?	SMURF::GREBUS	DTN 381-1426 - Digital Unix	`Wed Apr 16 1997 14:28`	20
	The statement that "all I/O is handled by CPU 0" isn't very accurate. It is true that interrupts are currently handled on one CPU, but that is only part (hopefully a small part) of the CPU cost of doing I/O. There is some work planned for the Steel release to distribute interrupts as well, but I don't know the details, and not all platforms and adapters may benefit. The 2100 and 4100 are no different in how the I/O generated CPU overhead is distributed. If this particular I/O workload generates a lot of interrupts (lots of small transfers) then I would expect it to work much better on a 4100 (which has much faster CPU's). There may be other factors too. 2100's tended to have SCSI adapters (NCR 810) which generated more CPU and memory bus overhead than later adapters. /gary
1170.4	What evidence?	WIBBIN::NOYCE	Pulling weeds, pickin' stones	`Thu Apr 17 1997 08:51`	8
	How did your customer reach the conclusion that his bad performance was "because all I/O is handled by CPU 0"? As others have mentioned, that's not mostly true, even on the 2100. If there was some other cause of the bad performance, attacking the wrong area won't help it. Also, what 2100 was being compared to what 4100? If you were comparing a 2100 4/200 to a 4100 5/400, I would expect the 4100 to be about 4x as fast...