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

Conference spezko::cluster

Title:+ OpenVMS Clusters - The best clusters in the world! +
Notice:This conference is COMPANY CONFIDENTIAL. See #1.3
Moderator:PROXY::MOORE
Created:Fri Aug 26 1988
Last Modified:Fri Jun 06 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:5320
Total number of notes:23384

5232.0. "CI Bottleneck due to HSJ50 ??" by PRSPSU::BOUGEANT (MCS Antony ...) Tue Feb 18 1997 11:46

         
	 Hi,

	 Since we've upgraded our HSJ42s to HSJ52s, we don't have attented
	performances (+300% I/Os up to 4600 I/O per seconde and +200% thruput
	up to 12,7 MB per seconde with a size of 4KB unit transfert). 

	 Configuration:

	      - OVMS V6.2
	      - Cluster of 6 AS8000-5/440 with 2 CIXCD on each (rev ucode 8)
	      - 2 Star Coupler
	      -	9 HSJ52 with 64MB cache (HSOF V5.0J-1)
	      - many RZ29s (rev 14) RZ28s (rev D442) and some EZ54 (rev 098)
		EZ64 (rev 064)
  
	  VIOC is disabled (for the moment) and about 90% of I/Os are Read
	 operations (with small I/Os size , about 5 KB). 
	  All units on HSJ have of MAXIMUM_CACHE_TRANSFERT_SIZE of 64 blocks
	 and WRITEBACK_CACHE enabled .

          Below are CI statistics when user response time problem occurs;
	at same time, on an electronic disk (EZ64) which is very used,
	QUEUE LENGTH groows from 2 to 20 and RESPONSE TIME from  20ms to 150ms
	(THRUPUT=1500MB,I/O rate =700).

	 A site audit seems to tell that CI Bandwidth is the bottleneck ?
	
	If I remember, CIXCD is able to use both paths A and B at same time
	only if remote node is not the same node. In fact, each path has a 
	bandwidth of 8,75MB (70Mb), 17,5MB for both. 
	 Below BLOCK TRANSFERTS KB THRUPUT reports for CI-0 7,2MB; I think
	that value is used only for data blocks , not for ACK/NOACK. But in 
	this case, we don't achive 17,5MB ?

	Has anybody an idea (at least about xmit/receive on both paths) ?


	Regards.
	Luc BOUGEANT.

                                                                       Block
 Cir-                      Path  Data-  Mess-    Disk       Disk      Transfers
 cuit    Node  Component   Name  grams   ages Operations KB Thruput  KB Thruput
------  ------ ---------   ----  -----  ----- ---------- ----------  ----------
CI-0           ** total **         0.0  20116     1965.1     7142.2     7263.0
CI-0    ARAGON UNKNOWN     PNA0    0.0 8736.0      210.3      646.3      654.4
CI-0    BARRES UNKNOWN     PNA0    0.0 3269.7      832.1     4251.4     4345.3
CI-0    GIDE   UNKNOWN     PNA0    0.0 7309.4      188.8      494.6      502.6
CI-0    HSJA00                     0.0  538.0      280.9     1328.1     1367.9
CI-0    HSJA01                     0.0  684.6      343.5      902.4      903.0
CI-0    HSJA02                     0.0  313.6      156.8      990.9      991.1
CI-0    HSJA03                     0.0  237.5      119.1      560.3      560.5
CI-0    HSJA04                     0.0  751.6      377.7     1056.6     1056.8
CI-0    HSJA05                     0.0  567.7      286.1     1331.1     1368.5
CI-0    HSJA06                     0.0  145.2       73.2      173.0      173.0
CI-0    HSJA07                     0.0  334.4      168.2      363.2      363.4
CI-0    HSJA08                     0.0  270.0      136.1      332.6      333.3
CI-0    HSJA09                     0.0   47.1       23.5      104.2      104.3
CI-0    LOTI   UNKNOWN     PNA0    0.0 8757.9      187.9      708.8      720.6
CI-0    MORAND UNKNOWN     PNA0    0.0 7330.7      390.5      793.4      810.6
CI-0    PROUST UNKNOWN     PNA0    0.0  939.1      155.6      247.7      270.6
CI-1           ** total **         0.0 3553.2     1631.8     6071.1     6211.7
CI-1    ARAGON UNKNOWN     PNB0    0.0  375.0      167.8      540.0      602.6
CI-1    BARRES UNKNOWN     PNB0    0.0 1475.9      733.7     3768.9     3810.7
CI-1    GIDE   UNKNOWN     PNB0    0.0  284.5      139.6      401.9      404.5
CI-1    HSJB00                     0.0  434.9      215.8      454.3      457.8
CI-1    HSJB01                     0.0   29.8       14.9       66.7       66.7
CI-1    HSJB02                     0.0 1052.3      508.8     1813.7     1918.0
CI-1    HSJB03                     0.0  343.7      171.3      643.0      645.2
CI-1    HSJB04                     0.0  387.6      192.9      591.0      595.2
CI-1    HSJB05                     0.0  307.5      153.5      982.1      982.7
CI-1    HSJB06                     0.0  336.9      168.3      832.6      833.2
CI-1    HSJB07                     0.0  413.1      206.1      687.9      688.8
CI-1    HSJB08                     0.0    0.1        0.0        0.0        0.0
CI-1    HSJB09                     0.0    0.1        0.0        0.0        0.0
CI-1    LOTI   UNKNOWN     PNB0    0.0  282.8      140.6      496.2      497.8
CI-1    MORAND UNKNOWN     PNB0    0.0  851.8      299.5      621.5      652.8
CI-1    PROUST UNKNOWN     PNB0    0.0  530.3      150.6      242.7      267.2

T.RTitleUserPersonal
Name
DateLines
5232.1Add more STAR'sZUR01::GLAUSFeel like Don Quijote...Wed Feb 19 1997 04:228
    Adding faster HSJ's just dont give you more performance when
    your lock and data traffic is that high already.
    
    Adding another star or two will give you much more performance gain,
    even better, go to 3-4 CIPCA's per Host and use the 4KB packet
    and arbitration option (not currently used in your config).
    
    Guido.