f 027 4503 4

HOW THE NEWMEMOREX 3680MEASURES UP.

A DP MANAGER'S GUIDETO DISC DRIVE

PERFORMANCE EVALUATIONS.

> I I "'Iu orrouaoprrni

slouboiq 9ldil6qmoo gulq ylleohotsiHb9on6(in9 n99d 9vbi1 x9iom9M yd b9qol9V9b.3m9Jaya gnifcix9 ot t09qa9i diiw yld6i9bisnoo

mlit nirit to tn9vbs erit ,i9V9wori ,abnim ynsm nl-muoiio to 198 W9n s b9t69io 8Bf1 ygolonrto9t

niiil gniqol9V9b to aeitluoittib 9rll n9vi0.39onst3ot IbojqoI Ji ai ,906lq teiit erii ni aeitilidBqso mlit

CMOS) i9iut06tunBM 9ldit6qmoO guH b t09qx9?gnh9fto Isilini lieril rltiw lis 9rit9on6vb6 oals ot

.ynBqmoo srii no abri9q9b tiylauoivdO9V6fi ,9lqmBX9 not ram9taya mlit niritx9iom9Mas bnA .YYGt 9onia tn9mqol9V9b i9bnu n99d

,am9taya bB9(i mlit nirlt to 89ii9a b ni bnoo93 9fit-i9bianoo 3to9lt9i yb69il6 0838 x9iom9M 9dt

.noitulov9touboiq9lds

i9tan6it Btsb ytiosqeo 9Q6iota to ami9t nl9f1t ot l6oitn9bi ai ti ,ab99qa 339006 bne 9tsi

noitiaoq ythsq sidt 9ono ri9V9woH .m9taya 08889dt to ai9ngi39b srlt rb9do69i asw 0888 9iit rltiw

atn9m9on6fin9 ot noiJn9tts ii9dt b9mut 0888yllBoitemeib tuqdguoirit 9voiqmi bluow tBiit

rdgid 916 tn9m9V9idoB ns doua lot 39>l6ta 9riT9V6d nso tn9m9voiqmi t89bom s n9V9 93U609d

.3193U lot 39on9up9snoo gnidsinotsB

CTM3TMO0 TO 3J8AT

t noitouboitnl

t luqdguoidt b9voiqmi iot 9360 9dT\ 3bodt9m 339006 Bteb leohotsiH

S 3bodt9m gnidtsq 08888 3bodt9m gnidteq 0888£ 0838 .sv 0888 :sno3h6qmoo >h6mriori983 b9>l36 n9tto tsom snoit39up 9dT8 ?9i9d mart og uoy ob 9i9dW

1111 1' i_LL ' • 'CONTENTS

>i >| i • I i I i i i I i i i I i i i IINTRODUCTION

i I i | i

• i

i | i j i | i I i | i | iTABLE OF

Introduction 1

The case for improved throughput 1Historical data access methods 1

3380 pathing methods 23680 pathing methods 3Benchmark comparisons: 3380 vs. 3680 4The questions most often asked 6Where do you go from here? 8

Historically, plug compatible productsdeveloped by Memorex have been enhancedconsiderably with respect to existing systems.In many minds, however, the advent of thin filmtechnology has created a new set of circumstances. Given the difficulties of developing thinfilm capabilities in the first place, is it logical toexpect a Plug Compatible Manufacturer CPCM)to also advance the art with their initial offering?

Obviously it depends on the company.Memorex thin film systems, for example, havebeen under development since 1977. And asthe second in a series of thin film head systems,the Memorex 3680 already reflects considerable product evolution.

In terms of storage capacity, data transferrate and access speeds, it is identical to the3380 system. However, once this parity positionwith the 3380 was reached, the designers of the3680 turned their attention to enhancementsthat would improve throughput dramatically.The stakes for such an achievement are high,because even a modest improvement can haveastonishing consequences for users.

The case for improved throughput

A study of interactive user productivity byIBM's Arvind Thadhani1 identified a significantincrease in productivity as system response timedecreased from five seconds to one second.Moreover, as system response time droppedinto the subsecond range, the increase in userproductivity skyrocketed Csee Figure 1). In fact,in the range between 3.0 seconds and 0.3seconds, user productivity increased almostthree times as fast as the system response timedecreased.

O 300

1 2SYSTEM RESPONSE TIME

(SECONDS)

Figure 1: Plot of System Response Timevs. User Transaction.

The implications in terms of bottom linebenefits are impossible to ignore. In one engineering study, for example, users working aton-line graphics terminals with a light penimproved productivity by as much as 550%when system response time decreased from1.5 seconds to 0.4 seconds.

In still another study involving code generation, a test project was completed in just 61%of the time projected when system responsetime was decreased from 2.22 seconds to 0.84seconds. What's more, there was a 57% reduction in the error rate.

Clearly if user productivity increases 2 to 5times, the cost savings quickly soar to significantproportions, but there are other ways to capitalize on throughput improvement. You could, forexample, add more terminals without degradingyour present system response time. And youcould obviously improve the turnaround timeon your batch processing. Scheduling also

; 1 '

improves because even peak loads can bedispatched within predictable and consistenttime frames. All told, the effect on DP operations is similar to adding a more powerfulprocessor... an expensive eventuality mostcompanies would prefer to defer as long aspossible.

Historical data storage accessarchitecture.

S 30

30 60 90 120 150 180

SUBSYSTEM I/O OPERATIONS PER SECOND

Figure 2: Plot of l/O's vs. Time.

As processing power dramatically increasedthe number of l/O's executed, accessing theburgeoning quantities of data continued to bepossible only at the expense of processingspeed. The fundamental problem was simplythat the l/O's couldn't get out of their own way.(Figure 2.)

Various access architectures were developedin an effort to improve the throughput. The mainobjective was to prevent a single I/O fromblocking access to the entire string as itobviously would do ifall the actuators were simplyconnected in series. Thus, there were solutionsthat provided multiple paths, various pathswitching arrangements, parallel-series hookuparrangements and so forth. Some workedbetter than others but all of them improved dataaccess.

In 3380 class subsystems, the sophisticationof the data access architecture is a constantreminder of the importance of the problem.

1UFASTER IS BETTER-A Business Case ForSubsecond Response Time',' COMPUTERWORLD,April 18,1983; Vol. 37, No. 16.

3380 PATHING METHODS. • | • i

i . I . i

The IBM 3380 Subsystem utilizes aconfiguration called Dynamic Path Selection(DPS) whereby groups of four actuators sharean internal path. (Figure 3.) The four internalpaths are connected to the two string controllersthrough a "Two by Four Switch" which allowsthe paths to be connected to either stringcontroller

In order to establish a connection betweena Storage Director, a String Controller and anactuator, it is necessary not only that these befree, but necessary also that no other actuatoron the same internal path be communicatingwith the other string controller. This additionalrequirement decreases the probability of accessand contributes to slower response becausecommunications with any one actuator effectively ties up 2,520 megabytes of data, thestorage capacity of the four actuators onthat path.

To minimize this problem, the 3380 isfrequently configured as two half strings. Thisimproves the overall response time comparedwith a full string because it reduces the contention for internal paths. This solution has adownside, however, in that each string, whethershort or long, requires an expensive stringcontroller. And for every two string controllers,an equally expensive storage controller isrequired.

i.I.i i.l.i

STORAGE DIRECTOR STORAGE DIRECTOR

. i

3880-3

STORAGE

CONTROL

STRING CONTROLLER STRING CONTROLLER

s: 3}2x4 SWITCH & 3380 AA4

ACTUATOR 0 ACTUATOR 2

ACTUATOR 1S

ACTUATOR 3S

£ACTUATOR 4

ftACTUATOR 6

ACTUATOR 5 ACTUATOR 7s 3380 B4

ACTUATOR 8 ACTUATOR 10

3380 B4

ACTUATOR 9 ACTUATOR 11

s: £ACTUATOR 12 ACTUATOR 14

ACTUATOR 13 ACTUATOR 15S 3380 B4

V

Figure 3: 3380 Pathing Method.

I I I I I I

3680 PATHING METHODS

STORAGE DIRECTOR STORAGE DIRECTOR

3888 DUAL

DIRECTOR

STORAGE

CONTROL

UNIT

STRING CONTROLLER STRING CONTROLLER )3683 DUAL

STRING

CONTROLLER

ACTUATOR 0

ACTUATOR 1s

3680 DISK MODULE

£ACTUATOR 2

3680 DISK MODULE

ACTUATOR 3

rJ

ACTUATOR 14

ACTUATOR 15S

3680 DISK MODULE

Figure 4: 3680 Pathing Method.

The Memorex 3680 Subsystem CFigure 4.)employs a pathing arrangement called "Maximum Availability Path Selection" or "MAPS"whereby each actuator is dual pathed and totallyindependent from each other. Together withseparate actuator microprocessors, this enablessimultaneous data transfer from any two actuators within a string, including to and from thesame HDA.

Since the two data paths have the ability toaccess all of the actuators within the string, thedata availabilityof the 3680 is far superior tothat of the 3380, whether it is configured as ahalf string or a full string. In fact, our DASDPerformance Analytical Model shows that a3680 full string will outperform a 3380 half stringsimply because the maximum unit of contentionin a 3680 string is only 630 megabytes —theamount of data stored under a single actuator.

i i i I i i i

i i i .• i i i.i i

This same model currently in use alsopredicts that MAPS will provide I/O transactionrates comparable to those of a 3380 configuredwith the Dynamic Reconnect feature ofMVS/XA. In fact, when 3680's are optimizedfor use with MVS/XA, our models predict theywill be able to use Dynamic Reconnect to evengreater advantage so that the improvementalready achieved will be even greater.

The MAPS feature also incorporates animproved protocol capability, which enablesmore data to be transferred per control operation. For example, MAPS transfers seek andset sector information in one control operationand does not need additional control operationsto verify the transfer. With the 3380 system,additional control operations are required toverify seek and set sector and to verify theparameters transferred from these first twooperations.

Further improvement in throughput resultsby staggering a track's index mark 180 degreesapart on adjacent cylinders. This speedssequential processing because less time isrequired for latency. The result is that data canbe retrieved faster because the system canstart looking for itsooner.

•

BENCHMARK COMPARISONS: THE 3680 vs. THE 3380

Described in the following are the initialbenchmark testing results which compare3680 subsystem performance with that of the3380 subsystem.

Figure 5 summarizes the results of abenchmark run for a large U.S. customer. Thebenchmark jobstream consisted of a programwhich executed job steps writing sequentialrecords using multiple blocksizes. The programthen did random reads using the same block-sizes and data sets created by the sequentialwrites.

Multiple jobstreams were run against multipleactuators on identical configuration 3680 and3380 subsystems. These jobstreams wereexecuted on a 3083, model E16 processorcomplex controlled by an MVS operatingsystem.

The 4x4 test demonstrates the comparisonfound on identical logical volumes when fourcopies of the jobstream were run against twoHDAs or four actuators on both the Memorex3680 and IBM 3380 configurations. In the 3380case, each HDA Cand thus its two actuators)was on a different Dynamic Path Selection[DPS) internal path, so the unit of contentionwas kept to the minimum 1260 megabytes.The 2x2 tests demonstrate the comparisonfound on identical logical volumes when twocopies of the jobstream were run againstone HDA.

The 4x4 test constitutes the best case for the3380, in that the HDAs were on different DPSinternal paths. However by using this configuration an improvement close to what may typically be expected is produced. The 2x2 testconstitutes the worse case for the 3380 in that

. i

the two actuators were on the same DPS internal path, while the Memorex case could fullyutilize the unrestricted dual paths providedby MAPS.

You can see a further example of the MAPSbenefit in Figure 6. It summarizes results of aprogram which provides the ability to select I/Otransaction rates, block sizes, read or writeoperations, and access modes. This was executed on the same processor with the samesoftware as the preceding benchmark, but onMemorex and IBM strings with 4 spindles and8 actuators.

PROGRAM THROUGHPUT IMPROVEMENT

(MRXTOIBM)

90

80

70

Oo

£ 60co

3 50

40

D 30niCOQ.

3 20HI

10

TIME

DIFFERENCE

A% FASTER

TEST

22:34 21:06 18:52 16:46

26% 24% 22%

4X4

• MEMOREX• IBM

20%

Figure 5: Benchmark Comparisons.

30:56 32:32

39% 40%

2x2

The results show that the throughput andperformance improvements are sustainedthroughout common (4K and 6K) and lesscommon [1K and 2K] blocksizes. They alsoshow the improvements increase as I/Orates grow.

The most apparent observation that maybe made with this benchmark is the Memorexsubsystem's ability to handle significantly higherI/O transaction rates. This will provide benefitsin peak transaction periods and when l/O's areskewed to a few actuators in the subsystem.It will also enable the Memorex subsystem tomaintain a more stable transaction time at aterminal or turnaround for batch workload.

For terminal response time this is particularlyimportant, as we have earlier discussed.

As you can see from these results, the3680 disc subsystem can significantly outperform the 3380 disc subsystem.

Performance comparisons are continuouslybeing made by Memorex; if you would like toreceive copies of future benchmark reports,just write "Benchmarks" on your business cardand mail it to us at the address on the back ofthe brochure.

I/O THROUGHPUT IMPROVEMENT

(MRX TO IBM)

AVERAGE r\INCREASEW

OoUJCO

ccLUQ-

200

180

160

140

120

100

80

60

40

20

0

(I/O RESPONSE)

BLOCK

SIZE

+42%

15 16 17

1K RANDOM

• MEMOREX• IBM

Figure 6: Benchmark Comparisons.

16

+35% +40% +42%

n

17 18 19 19 20 21 16 26

2K RANDOM 4K RANDOM 6K SEQUENTIAL

1 I ' i ' I ' I I l_ ' I ' i • I I I ' I I I I I IE QUESTIONS MOST OFTEN ASKED

Q: Do Ineed additional or special software forMAPS?

A: No. MAPS istransparent to the operatingsystem and will be supported by any operatingsystem supporting 3380 subsystems.

Q: What is Dual Path?A: Dual Path is a feature which provides a

second path to each actuator within thestring, permitting simultaneous read/writeoperations on any two different actuatorswithin the string -even on the same HeadDisc Assembly [HDA).

Q: How does it improve throughputperformance?

A: Each of the two data paths can control anyactuator in the string by means of theMaximum Availability Path Selection [MAPS).This feature improves data availability byreducing the Unit of Contention, FalseDevice or Actuator Busy and missed RPSinterrupts. Throughput is further facilitatedby an improved protocol that speeds communications between components ofthe system.

Q: What is the Unit of Contention?A: The amount of data that becomes unavail

able due to blocking by a data transfer atany one actuator. With the 3380 2x4 switch,one actuator can block out 3 others at 630MB each for a total of 2,520 MB. On the3680, the Unit of Contention is only 630 MB[the actuator in use) because alternate datapaths solve the blocking problem.

6

i i i I i i i I | ! | I ! I I I I I j I I I I I I I

1 I I I . I I

Q: What is False Device Busy?A: False Device or Actuator Busy is a condition

where a component constituting the path tothe device or actuator is busy and cannot berecognized by the I/O control software. Inthis situation, itfalsely appears that the deviceor actuator is busy. If, for example, an internalpath of a 3380 is busy because a particularactuator is in use, attempts to utilize otheractuators on the same path will result in afalse busy.

Q: What is a Missed RPS Interrupt?A: A Rotational Position Sensing Interrupt

signals that a desired data location is aboutto pass beneath the read/write head. Ifthedata location cannot be accessed due tothe blocking problem discussed earlier, aMissed RPS Interrupt occurs. This introduces a delay of at least 16.6 ms which isthe time required for an additional 360°rotation of the disc.

Q: What is Dynamic Reconnect?A: Dynamic Reconnect is a feature of the

MVS/XA operating system which uses hardware in the channel subsystem to providea different reconnect path back to theprocessor.

Q: Will MAPS support Dynamic Reconnect?A: Yes. Current 3680's offer comparable

throughput and performance to 3380's withDynamic Reconnect; when Memorex 3680Dynamic Reconnect is implemented, theimprovement will be dramatic.

.1.1.1

'' I ' *

1 . I . I

Q: What is a 3380 internal path?A: A path within a 3380 string to which up to

four actuators may be attached serially.

Q: What is Extended Architecture?A: Extended Architecture (XA) is a new hard

ware architecture developed by IBM thatprovides 31-bit addressing and moved I/Oprocessing from the CPU to an ExternalData Controller CEXDC).

Q: Is there any internal path queueing?A: No. I/O queues are not formed for the

internal paths of a 3380 string. Should an I/Obe issued for an actuator on an internal pathalready busy, a false busy will occur.

Q: Does MAPS alter the ratio of 3680's requiredto replace 3350's?

A: Yes. For the 3380, IBM at times has recommended a conversion ratio of 1.5 to 1;however, the 3680 ratio is 2 to 1. This is dueto the throughput improvement of the 3680which permits greater effective use of thestorage capacity.

Q: Does block size affect the performance ofMAPS?

A: Yes. The improved protocol of MAPSreduces the overhead associated with everyI/O operation. Therefore should less effectivesmaller block sizes have to be used with3380 technology, the 3680 benefitincrease.

I.I.I. .1.1.1 . i . I . I

Q: Is path utilization affected by MAPS?A: Yes. Due to the enhanced protocol, utiliza

tion of components that constitute a path isreduced. Therefore, for the same amount ofpath utilization, the number of l/O's will begreater.

Q: What channel queueing algorithm shouldI use for MAPS?

A: To insure both paths are utilized effectivelywith a single processor, the Channel RotateBalanced Scheduling algorithm should beused. When multiple processors areattached to the 3888, then the Last ChannelUsed Scheduling algorithm should be used.

Q: Do other PCM's have the equivalent ofMAPS?

A: Not really. Other PCM's offer Dual Portcapability which share similarities withMAPS, however the improved protocolingof MAPS together with staggered indexmarks improves the throughput potentialconsiderably.

7

I I I

I I I I I I I II I I I I I I

THE NEXT STEPi i i

• i i i » • i i I i i i

• i i

... is to prove it to yourself. In addition tothe conclusions reached via benchmarks, it ispossible to predict with considerable accuracythe improvement in throughput performancefor any given DP operation. Using an advancedanalytical model, Memorex Systems Engineerscan perform on-site evaluations and provideyou with documented projections, typically inone day.

An even more exacting evaluation can beperformed in Santa Clara, California, usingyour software to develop comparisons on both3380 and 3680 subsystems.

We invite you to be amazed.

i i i I i i i i « | • i •

i.i.i.

i » I > i •

i.i.i.

i •

I I I I I 1 I '

q3T8TX3M3KT

ot noitibbs nl .tl98iuoy ot ti 9voiq ot ai...ai ti (8>h6mdon9d Bivbedos9i anoiaulonoo 9rit

yosiuooB 9ld6i9bianoo dtiw totaq ot 9ldi8aoq9on6mioh9q JuqdQuoirlJ ni tn9m9voiqmi 9iit

D9onBvbs ns gniaU .noitBi9qo SQ n9vig yns not2i99nign3 amgtayS x9iom9M ,l9bom IsoitylsnE

9bivoiq bns 8noit6ul6V9 9tia-no rmoh9q nsoni yllsoiqyt ,anoito9[oiq betn9muoob rltiw uoy

.ysb 9no

9d nso noit6ulev9 gnit06X9 9iom n9V9 nApniau ,6imotil60,bibIO BtnsS ni b9mioti9q

rlfod no anoahsqmoo qol9V9b ot 9iswttoa iuoy.ametayadua 0888 bns 0888

.bexsms 9d ot uoy 9tivni 9W

Memorex Corporation

San Tomas at Central ExpresswaySanta Clara, California 95052Telephone: (408) 987-2301

When it matters, make it Memorex

MEMOREX

"'/If:?/. 7Xut^v\V/

ktiiw

///n