Conference ulysse::rdb_vms_competition

1339.0. "Rdb on AXP CLuster: 3692 TPS-A" by NOVA::BANERJEE () Tue Apr 12 1994 00:19

From:	NOVA::HAGAN "Steve 381-2425 Engr Mgr INDUSTRY LEADING Rdb + DBMS  08-Apr-1994 1931"  8-APR-1994 19:37:14.51
To:	@PALMER,@RDBNEWS.DIS
CC:	HAGAN
Subj:	RDB + ALPHA DOES 3,692 TPS-A; TRIPLE THE WORLD RECORD

+-+-+-+-+-+-+-+
|d i g i t a l|         I N T E R O F F I C E   M E M O R A N D U M
+-+-+-+-+-+-+-+		Digital Confidential, Internal Use Only

TO: Bob Palmer 
	     				DATE:  April 8, 1994
                                        FROM:  Steve Hagan
                                        DEPT:  Rdb/DBMS Engineering Mgr
					EXT:   381 - 2425
        				LOC:   ZK2-1/O.C.1 
        				ENET:  NOVA::HAGAN
       
cc: Distribution

SUBJ: ALPHA + RDB = 	3,692 TPS-A	 TRIPLES THE WORLD RECORD 

1) HIGHLIGHTS, 2) MARKET POSITIONING, 3) CONFIGURATION 4) KUDOS

This week, we successfully audited 3,692 TPS and $4,873 / TPS 
on an ALPHA CLUSTER, VMS, Rdb, ACMS, Digital Compilers.
These results are ready for public release.

This is TRIPLE the benchmark ever run by a relational database on any platform.

Cluster was 4 - Alpha 7650s.

This proves the SPEED and SCALABILITY and AVAILABILITY of the ALPHA platform, 
both in SMP and in a Cluster. The database size was 3/4 TERABYTE
(766 gigabytes), with stringent ACID (failure scenario) processing required.

This benchmark was run in CLIENT/SERVER mode, serving 36,920 users.
                      
IN ADDITION:

Rdb + ALPHA hold the FOUR best price-performance records in the industry:

        System                                  TPS A     $/TPS A

 Digital 2100-A500MP 4 CPU Client/Server        662.32   $4,401

 Digital 2100-A500MP 1 CPU Client/Server        265.03   $4,405
 
 Digital DEC 4000-720 AXP 2 CPU C/S		402.76	 $4,861
 
 Digital DEC 7650 AXP 4 nodes		       3692.67	 $4,873

--------------------------------------------------------------------------

The message to IBM mainframe customers: 
	UPSIZE your capabilities while DOWNSIZING your costs.
        
To HP + SUN + RS6000: You are not in scalability game: 
	NONE of you can achieve these high end results.

To HP + SUN + RS6000 + Compaq: 
	You are NOT in the scalability game: 
	NONE of you can achieve these price / performance results,
	NOR the high end results.

These OLTP performance results combine with Rdb's WORLD RECORDS in
SORT and BACKUP/RESTORE, and the best high-availability + multimedia
capabilities in the industry, to make the DIGITAL solution the BEST wherever:

	Performance, Very Large Databases, High Availability, and Multimedia

are the requirements. This benchmark used a 240 disk database, yet we 
constructed and loaded the database in 6 1/2 hours.

NEXT BENCHMARKS: 	
	Rdb will set WORLD RECORDS in TPC-C (2 months).
	Rdb will show a 64 BIT GIANT MAIN MEMORY DATABASE DEMO, 
		proving the merits of a 64 bit chip.

For Internationalization markets:
	Rdb 6.0 shipped in Japan 27 DAYS after shipping in the U.S.
	A WORLD RECORD; Rdb is ISO SQL 92 Compliant, implementing ISO9000.

For the OBJECT ORIENTED Markets: 
	Rdb 6.2 will allow uses/customers to add Object Oriented Class 
	Libraries to Rdb databases. 
	TEXT oriented Content Based Retrieval will be the first class library;
	OLE support will also be in Rdb 6.2.
	NT operating system will also be supported.


2) MARKET POSITION SUGGESTIONS

	IBM Downsizing

		Alpha platforms, due to SMP + CLUSTER Scaling, 
		can OUTPERFORM the largest mainframe, and do it
		at a better price.
		IBM has a TPC-A number of 3509 TPS, $7.9 K/TPS.
		Counter: this was run with TPF (Transaction Processing 
		Facility): also known as SABRE - Airline Reservation System
		This is a special purpose, no O/S, custom package.
		MESSAGE: Digital, with a general purpose relational database,
		good tools, general purpose operating system, and scaleable 
		platform, can provide your customer with more capability 
		than even the famous IBM airline reservation, with their 
		largest mainframe + I/O technology.

	PC Desktop UPSIZING

		Digital, at the low end, has better price performance
		than Compaq. The advantage is that a Digital solution 
		can then scale up to larger than an IBM mainframe for 
		capacity. No other vendor can do this.

	HP, SUN, IBM RS6000
	
		None of these vendors have a cluster scaling story 
		to compare with the Digital story. ASK for their benchmark;
		(they don't have one). Then explain to the customer 
		that an OFFICIAL TPC-A requires demonstrating 
		failure and recovery from losing a disk, a CPU, 
		an entire NODE in a cluster or the entire cluster. 
		NO ONE ELSE CAN DO THIS TODAY! DIGITAL HAS IT NOW!


APPROXIMATE CONFIGURATION:

The precise system is described in the TPC report; available 
from the author on request.

	The cluster benchmark test not only validates the capability of
	Alpha AXP, OpenVMS and Rdb to handle extremely large throughput volumes,
	but also the ability to support very large databases on very 
	large hardware configurations. The various ACID tests demonstrate that
	this system can recover this huge database correctly from the 
	failure of a disk, a node and even the entire cluster.

	Attached are some vital information on the benchmark.

	Cluster Test Database Size
	--------------------------
	TPC-A Result = 3692.02 TPS-A at $4873/TPS
	
	Number of concurrent users		      36960

	Number of disks (with 10% spare)		303

	Total size of database				747 GB
		History table size	638 GB
		Account table size	 53 GB
		After-image journal	 57 GB
		Other misc		  1 GB

	Number of records in Account table		369 Million

	Important Throughput Data
	-------------------------
	
	106 Million transactions per (8-hour) day
	
	8122 disk I/Os every second

	7384 network I/Os every second (messages between front and back-end)

	Cluster Test Hardware Configuration
	-----------------------------------

	Back-end System
		4 node DEC 7650 AXP (200 MHz) cluster
	        512 MB memory per each node
		303 disks
		22 HSJ40 SCSI controllers
		4 TZ87 (20GB) tape drives
	        FDDI connection

	Frond-end System
		44 MicroVAX 3100/90 with 32 MB Memory on each
		44 disks 

	Terminals
		36960 VT510 terminals
		 5082 DECserver 90L+ terminal servers (with 10% Spare)

	Cluster Test Software
	---------------------

	Back-end System
		OpenVMS AXP V6.1 
		DEC Rdb V6.1
		DECnet
		Cluster Software

	Front-end System
		OpenVMS VAX V5.5 
		ACMS V3.3
		Rdb V6.1
		VAX C
		VAX CDD+
		DECnet

NOTES:
	The scaling in the SMPs and in particular the cluster was 
	better than expected.
	We actually ran over 4,000 TPS in the server, and found no 
	bottlenecks in Rdb. We believe Rdb could do over 10,000 TPS-A,
	with enough hardware. 
	The audited benchmark had CPUs at only 90% busy, and we had to slow
	the test to stay within the planned TPS level. This is because
	we had a limited amount of disks, and were trying to audit by
	a deadline, April 8 (we made it).


---------------------------------------------------------------------------

THANK YOU's

These results, and their potential positive effects on Alpha Positioning
would not have been possible without the following people.

Bill Demmer, Pauline Nist, Steve Holmes

	They have loaned Rdb hardware for use in debugging + tuning,
	and sponsored the entire benchmark. Without their faith and 
	support, this effort would never have gotten off the ground.

Charlie Christ, 
	
	He + his people have loaned us an enormous disk configuration.

Sas Durvasula

	Sas + I went far out on a limb for this one. Sas continued to
	defend this effort, even through our darkest days.

Bhagyam Moses, Chris Janson, Lee Allison, Joe McFadden
	Ruth Morgenstein, Zarka Cvetanovich, Bob Sibley	

	All kept pushing, finding new hardware + resources as we exceeded 
	plans, replacing where needed, helping schedule + plan through long
	nights + days.

Storage: Ritchie Lary

	Ritchie worked extensively with the engineers to solve I/O 
	bandwidth issues, how to optimally configure the systems, 
	and was instrumental in making hardware and microcode available 
	as needed.

        Keith Parris, Dave Clark, Bob Passmore, Doug Williams	
	Host based RAID help, Arranged firmware upgrades, 
	Arranged disks, Loaned disks for 3k & Sable tests

VMS: Drew Mason, Paul Houlihan, Ray Pfau, Phil Norwich,Brad Waters	
	Buffer objects, Critical bugfix with system crashes
	Consulting, IOX images

Walt Kohler, Wael bahaa-El-Din, Mike Brey, Yongmin Chen, Paula Deangelis 

	Understanding the TPC requirements, modeling Rdb on the several
	configuration, recommendations on where to tune, and helping
	with the Sable tests; these all fed the final results.

Jim Gray

	The world leader in transaction processing, the inventor 
	of the TPC benchmark itself: the seeds you planted have become 
	the greatest Sequoia in the woods. You were the knowledgeable 
	mentor that everyone went to when ideas needed reviewed, encouragement 
	was needed, sanity needed restored. Thank You.

Peter Spiro, Rick Anderson, Arun Gopolan

	All were involved in new upgrades to Rdb and a high pace, to make 
	this doable.

DBS Systems group: Kevin Longfellow & Mike Taylor	
------------------------------------------------------------------------------
THE HARD CORE FOUR
These are the four people who put in the time, blood, sweat, and tears
to make this landmark happen.

Dave Walrath, Gigi Lirot 
	Worked intensly and actively as part of the engineering team for months.
	Thank You from everyone.

RABAH MEDIOUNI and T.K. RENGARAJIAN.
	
	These two Rdb engineers were the heart and soul of the benchmark.
	They worked long days, nights, weekends for MONTHS making this happen.
	New ideas flowed forth constantly - no obstacle was taken as final
	no problem was too large or too small; all were resolved.

Rdb Engineering: Everyone.

	This is a new landmark in the world of computing. 
	It is rare in a professional engineers life when they can say
	that they were the best, the best in the entire world, 
	beyond any doubt.
	You have done this, by working together as a team, focused on 
	the shared goal of building the best product in the world.

	My congratulations to you all, savor the day.

Steve Hagan
Rdb Engineering Manager