Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 610 (1995)

(USC DC Other)

USC Computer Science Technical Reports, no. 610 (1995)

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content T ec hniques to Quan tify SCSI Disk Subsystem Sp ecications for
Multimedia
Shahram Gandeharizadeh James Stone Roger Zimm erm ann
Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles California Marc h Abstract
Magnetic disk tec hnology has established itself as the mass storage device of c hoice in the commercial
arena Most often the op erating system of a hardw are platform that emplo ys this device eg a
Unix
TM
based w orkstation hides its ph ysical attributes b y conceptualizing it as an arra y of blo c ks
This abstraction exp edites program dev elopmen t time b ecause the programmer is no longer concerned
with the w orking details of the mass storage device Moreo v er the nal program is p ortable as long as
a new target platform pro vides an iden tical abstraction of its magnetic disk driv e
This paradigm is eectiv e for those applications with no realtime constrain ts Ho w ev er certain
applications eg m ultimedia cannot tolerate signican tv ariations in the service time of the disk driv e
They m ust estimate the service time of the disk accurately in order to b oth supp ort realtime constrain ts
of the application and sc hedule the disk eectiv ely in the presence of m ultiple requests Assuming a Unix
based w orkstation this study rep orts on our exp erimen tal tec hniques to iden tify the ph ysical details of
a magnetic disk driv e The mo dels constructed based on the results obtained from these exp erimen ts
can b e used b y a program to satisfy its realtime constrain ts
In tro duction
The magnetic disk driv e tec hnology has b eneted from more than t w o decades of researc h and dev elopmen t
It has ev olv ed to pro vide a lo w latency in the order of milliseconds and a lo w cost p er megab yte of storage
appro ximately cen ts at the time of this writing It has b ecome common place with ann ual sales in
excess of billion dollars oST far exceeding that of b oth CD Driv es and T ap e recorders whose
com bined sales is less than billion dollars Both w orkstations and p ersonal computers emplo y magnetic
This researchw as supp orted in part b y the National Science F oundation under gran ts IRI IRI NYI
a w ard and CD A a gran t from DODIn tellig ence Comm unit y and a HewlettP ac k ard unrestricted cashequipmen t
gift
disks as nonv olatile storage devices Their op erating systems pro vide the application programmer with
the abstract p ersp ectiv e that the storage capacit y of this device is a linear arra y of xed size blo c ks Suc h
a system simplies program dev elopmentbyallo wing the programmer to fo cus on the task at hand instead
of the ph ysical attributes of the disk driv e Nonetheless a n umberofdata in tensiv e applications eg
m ultimedia m ust estimate the service time of a disk in order to satisfy their realtime constrain ts This
study fo cuses on the Unix op erating system due to its p opularit y and in v estigates tec hniques to iden tify
the ph ysical c haracteristics of a magnetic disk in order to estimate its service time These estimates can b e
used to dev elop implemen t and ev aluate a realtime application Our tec hniques emplo y no proprietary
information and are quite successful in dev eloping a metho dology to estimate the service time of a disk
Ho w ev er these tec hniques do iden tify a few attributes of magnetic disks that are dicult to explain
without proprietary information
Our prop osed tec hniques are a con tribution b ecause they demonstrate that a programmer do es not
need to abandon the b enets of the dev elopmenten vironmentpro vided b y an op erating system to supp ort
m ultimedia applications or other applications with a realtime constrain t A program is p ortable as long
as it consumes the ph ysical w orkings of a magnetic disk driv e as its input parameters Eac h time a program
is p orted to a new platform the system administrator computes the ph ysical c haracteristics of the new
disk driv e using our tec hniques and pro vides them as input to the program
T ec hniques to estimate the disk service time ha v e b een prop osed in R W Wil The previous studies
prop osed mo dels that pro vide an a v erage estimate of the disk service time These mo dels w ere targeted
for use in sim ulation studies to yield realistic results Our mo dels are designed to estimate the service time
of a disk driv e to enable an application program to satisfy its realtime constrain ts As suc h our mo dels
do not underestimate the service time of a disk driv e This ma y reduce the o v erall pro cessing capabilit y
throughput of a system ho w ev er this limitation can b e eliminated once a program is debugged b y
in tro ducing a sp ecialized soft w are la y er that tailors the system to the ph ysical details of a disk driv e

Moreo v er w e rep ort on the scalabilityc haracteristics of the SCSI bus This information has not b een
rep orted b y the previous literature
The rest of this pap er is organized as follo ws Section pro vides an o v erview of a target platform
congured with the Unix op erating system that emplo ys a magnetic disk driv e Next Section fo cuses
on m ultimedia information systems and describ es wh y it is imp ortan t to estimate the service time of a
disk driv e to maximize the n um b ers of sim ultaneous displa ys of video clips supp orted b y the system In
Sections and w e fo cus on the HPUX op erating system and dev elop sev eral exp erimen tal tec hniques
to iden tify the ph ysical c haracteristics of the disk driv e Using this information Section dev elops a
metho dology to estimate the service time of the disk driv e W e presen t brief conclusions and future
researc h directions in Section Ov erview of the T arget Dev elopmentEn vironmen t
SCSI-2
Adapter
Memory CPU
System bus
SCSI-2 bus
Disk Disk Disk
Initiator(s)
Target(s)
The Host adapter links
the system bus with the SCSI-2 bus.
Disk Subsystem
SCSI ID 0 SCSI ID 1 SCSI ID 15*
SCSI ID 7
*Note: Narrow SCSI supports 8 devices, while wide SCSI supports either 16 or 32 devices.
Figure Arc hitecture of the T arget En vironmen t
An abstract description of our target en vironmen t is pro vided in Figure The system bus pro vides a v ery
high p erformance nanosecond latency and more than MBytes p er second transfer rate once the bus
arbitration o v erhead is considered The SCSI Small Computer System In terface bus supp orts a p eak

Spindle
Head
Arm
Platter
Cylinder
Arm
assembly
Figure A disk driv e
transfer rate of or MBytes p er second dep ending on whether it is Narro w F ast or F astWide
resp ectiv ely The disk subsystem is cen tral to this study and is detailed in Section Section pro vides
an o v erview of our target op erating platform
Ov erview of the SCSI Disk Subsystem
This section starts with an o v erview of a magnetic disk driv e Subsequen tly it describ es the SCSI host
bus adapter and howit in terfaces with m ultiple disk driv es
Ov erview of a Magnetic Disk Driv e
A magnetic disk driveisa mec hanical device op erated b y its con trolling electronics The mec hanical
parts of the device consist of a stac k of platters that rotate in unison on a cen tral spindle see R W for
details Presen tly a single disk con tains one t w o or as man y as sixteen platters see Figure Eac h
platter surface has an asso ciated disk head resp onsible for reading and writing data Eac h platter is set
up to store data in a series of trac ks A single stac k of trac ks at a common distance from the spindle is
termed a cylinder T o access the data stored in a trac k the disk head m ust be positioned o v er it The
op eration to rep osition the head from the curren t trac k to the desired trac k is termed se ek Next the disk
m ust w ait for the desired data to rotate under the head This time is termed rotational latency time

T o meet the demands for a higher storage capacit y disk driv eman ufacturers ha vein tro duced disks with
zones A zone is an con tiguous collection of disk cylinders whose trac ks ha v e the same storage capacit y T rac ks are longer to w ards the outer p ortions of a disk platter as compared to the inner p ortions more data
ma y b e recorded in the outer trac ks While zoning increases the storage capacit y of the disk it pro duces
a disk that do es not ha v e a single transfer rate The m ultiple transfer rates are due to the v ariable
storage capacit y of the trac ks and a xed n umberofrev olutions p er second for the platters
Ov erview of the SCSI Bus Op eration
The SCSI standard is b oth a bus sp ecication and a command set to ecien tly use that bus see Ded Figure sho ws ho w the SCSI host bus adapter also called initiator links the system bus based on an y
standard or proprietary bus to the SCSI bus The storage devices also called tar gets are attac hed to
the SCSI bus The goal of the SCSI standard w as to pro vide a device indep enden t abstraction for dieren t
storage devices to the system F or example the only c haracteristic that dieren tiates among magnetic disk
driv es of dieren tv endors is their maxim um capacit y The SCSI devices and host bus adapters in our systems are of t wov arieties F ast MBytessecond
transfer rate and F ast Wide MBytessecond transfer rate The data path of a fast SCSI bus is one
b yte wide and it can supp ort a maxim um of eigh t devices due to its addressing sc heme One of them is the
host bus adapter itself Ev ery SCSI device has a xed priorit y that directly corresp onds to its SCSI ID
This n um b er ranges from to in v alue The host bus adapter usually is assigned ID n um b er b ecause it
corresp onds to the highest priorit y
A fast wide SCSI bus can transfer t wob ytes of data in parallel
and supp ort up to sixteen devices n um b ered to Because of the wider data path the bus bandwidth
is t wice that of the narrowv ersion
The transfer rate of a t ypical magnetic disk is to MBytes p er second The higher bandwidth of a
The wide SCSI sp ecication w as in tro duced after the narrowv ersion Consequen tly the priorities in terms of SCSI IDs
are with b eing the highest and the lo w est

SCSI bus
Disk 1
mechanism
Disk 2
mechanism
Disk 3
mechanism
Host sends
command
Controller
decodes it
Seek Rotational
latency
Data transfer off mechanism
SCSI bus data
transfers to host Status messages to host
Read service time for disk 1
Time
Read service time for disk 2
Figure Timing of o v erlapp ed read op erations with SCSI devices that implemen t the disconnectreconnect
feature
SCSI bus can b e utilized b ecause man y disks implemen t the disc onne ctr e c onne ct feature Figure sho ws
ho w the read op erations of sev eral disks can b e o v erlapp ed After receiving a command from the host bus
adapter the target disconnects itself from the bus and p erforms the seek op eration oine It then w aits
un til the rst of the requested blo c ks rotates underneath the disk head When the actual read op eration
starts the data is not immediately sen to v er the SCSI bus Instead it is stored in a small buer Whenev er
this buer is full the disk reconnects to the bus and transfers the accum ulated data to the host bus adapter
in a burst that matc hes the maxim um sp eed of the SCSI bus Hence a device that executes a long read
ma y b e connected to the bus for only to of the time This lea v es ro om for other devices on the
same bus to receiv e and execute commands concurren tly With this tec hnique of o v erlapp ed reads and
writes the o v erall throughput of the system can b e impro v ed Ho w ev er the SCSI bus starts to b ecome a
b ottleneckbey ond a certain n um b er of disks This causes requests referencing the lo w est priorit y disk to
observ e the w orst service time b ecause disks gain access to the bus based on priorit y Ov erview of the Unix Op erating System
Unix of whic h HPUX is a particular implemen tation is a generalpurp ose op erating system Sev eral of
its features are cen tral to this study These features are designed to mak e general system programming
and op eration easier Ho w ev er at times they hinder our exp erimen ts with a magnetic disk driv e

The access time of a mec hanical disk driveis m uchslo w er than that of system memory Therefore
Unix main tains the most recen tly referenced pages in a memory structure termed a buer p o ol in order
to minimize the n um b er of accesses to the disk driv e If a referenced page is residen t in the buer p o ol
no IO request is issued to the disk While this is a useful service its presence impacts our exp erimen ts
b ecause it migh t translate a disk page request to a memory reference
Unix pro vides a consistentin terface to devices All devices are abstracted as les whic hmaybe
op ened read from written to and closed using standard programming language functions An addition to
the op erating system termed a device driv er is pro vided for eac h sp ecic device to presen t this abstraction
Access to these device les uses one of t w o metho ds either through a blo ck in terface whic h emplo ys the
buer p o ol or through a r aw in terface whic hb ypasses the buer p o ol
The Unix system is b oth multitasking more than one pro cess ma y b e sim ultaneously executing at a
time and multiuser more than one user ma ybesim ultaneously executing pro cesses In addition to user
pro cesses the op erating system has pro cesses of its o wn that run without direct user in terv en tion These
system pro cesses p erio dically p erform v arious housek eeping tasks for the system and cannot b e preempted
b y user pro cesses
Con tin uous Displa y of Video Ob jects Using a Magnetic Disk Driv e
Video ob jects m ust b e retriev ed and displa y ed at a xed rate in order to ensure their con tin uous displa y Otherwise their displa y will suer from frequen t disruptions and dela ys termed hiccups T o simplify the
discussion assume that the bandwidth required to displa yan object R
C
is lo w er than the bandwidth of
the disk driv e T o supp ort a con tin uous displa y of a video ob ject X using a magnetic disk driv e sev eral
studies CL P ol BGMJ ha v e prop osed the follo wing approac h First ob ject X is partitioned in to
n sub ob jects X
X
X
n
The size of eac h sub ob ject is a function of the bandwidth required to ensure
a con tin uous displa y and the bandwidth of the disk driv e This size is xed for ob jects with the same

bandwidth requiremen t When an ob ject X is referenced the system stages its rst blo c k sa y X
in
memory and initiates its displa y Before the displa yof X
completes the system retriev es X
in to memory
in order to ensure a con tin uous displa y This pro cess is rep eated un til all blo c ks of an ob ject ha v e b een
displa y ed
Display W Display W
W W
Disk
Activity
System
Activity
W
i i+1 i+2
i
i+1
X
X
j j+1
Display X
j
T
wseek
Time Period (Tp)
Z
k
Z
k+1
Figure Time P erio d
Since the bandwidth of the disk exceeds R
C
the system can supp ort sim ultaneous displa ys of sev eral
ob jects Figure demonstrates the con tin uous displayof n ob jects In this gure the time to displa y
a blo c k is termed a time p erio d This time is partitioned in to slots with eac h slot corresp onding to the
retriev al time of a blo c k from the disk driv e The n um b er of slots in a time p erio d denes the n um ber
of sim ultaneous displa ys that can b e supp orted b y the system F or example a blo c k size of KBytes
corresp onding to a MPEG compressed mo vie R
C
Mbps has a second displa y time T
p
Assuming a magnetic disk with a transfer rate of Mbps tf r Mbps and maxim um seektimeof milliseconds suc h blo c ks can b e retriev ed in seconds Hence a single disk supp orts sim ultaneous
displa ys As demonstrated b y this example in order to maximize the n um b er of sim ultaneous displa ys
the system m ust accurately estimate the w orst seek time and transfer rate of the disk in order to sc hedule
requests eectiv ely
The previous example is a simplication b ecause it assumes a constan t transfer rate for the disk As
describ ed in Section the platters of curren tly a v ailable disk driv es are partitioned in to zones that
ha vea v arying n um b er of sectors This results in dieren t transfer rates for dieren t zones The system
designer should tak e the zoning information in to accoun t to b est utilize a disk driv e for the retriev al of
con tin uous media ob jects Otherwise the designer m ust assume the w orst case transfer rate that will result
in a considerable w aste of the disk bandwidth for all but one the slo w est zone
If the disk bandwidth is m ultiplexed among dieren t displa ys see Figure the disk heads m ust b e
rep ositioned to a dieren t lo cation for eac h sub ob ject This results in a w asteful seek op eration that
consumes the time T
w seek
see Figure y et transfers no data This time is a function of the distance in
cylinders that the disk head m ust tra v el The designer can enhance the utilization of the disk driv eb y
estimating the seek time accurately as opp osed to using the w orst case v alue
In addition to ensuring a con tin uous displa y a system designer m ust consider the p erformance require
men ts of an application The bandwidth of a single disk driv emigh t b e insucien t to supp ort the n um ber
of sim ultaneous displa ys required b y an application One ma y assume a m ultidisk platform consisting of
D disks to resolv e this limitation Ho w ev er notice from Figure that the SCSI bus is a shared resource
Increasing the v alue of D will not increase the n um ber of sim ultaneous displa ys when the SCSI bus
b ecomes a b ottlenec k resource See the discussion of Section F or this study w e underto ok exp erimen ts to learn ab out the SCSI disk IO subsystem the SCSI
bus zoning and the seek prole of sev eral SCSI disk driv es Ho w ev er all the disks a v ailable to us
ha v e cac he memory on the con troller b oard Although the cac he op eration normally only impro v es the
p erformance of a disk driv e w ew an ted to b e sure ab out its impact on our measuremen ts so w e devised
aw a y to selectiv ely enable and disable it The eect of the cac he op eration is men tioned in the follo wing
sections whenev er it is relev an t to the results

T est Platforms
T able lists the conguration of the platforms that w ere used for the tests in this study System System System T yp e W orkstation HP W orkstation HP CPU P ARISC MHz P ARISC MHz
Main Memory MB MB
SCSI Host Adapter Builtin F ast Builtin F ast Wide
SCSI System Disk Quan tum PDS MB HewlettP ac k ard C GB
SCSI Auxiliary Disks HewlettP ac k ard C GB Seagate STW GB
Op erating System HPUX HPUX C Compiler HPUX cc GNU gcc HPUX cc GNU gcc
Ph ysical Net w ork Connection Y es No
Op eration Mo de Singleuser Singleuser
T able Conguration of test platforms
HPUX Lo wlev el SCSI Programming In terface
In order to analyze the c haracteristics of a disk drivew e disabled the Unix buer manager to ensure that
all the read and write requests in our test programs w ould b e executed b y the ph ysical disk devices In
addition w e issued seek commands to the magnetic disks as w ell as enabled or disabled the ondevice read
cac hes In the follo wing w e describ e ho w this w as ac hiev ed using the HPUX op erating system
Bypassing InputOutput Buering
Unix do es not buer IO requests to a disk if it is accessed as a r aw device In order to use a device in
ra w mo de its corresp onding c haracter sp ecial le m ust b e op ened F or example in HPUX these les are
lo cated in the devrdsk directory and migh t app ear as
devrdskcds
After a disk is op ened as a ra w device the standard C library calls read and write can b e used to
access the data on the device Note ho w ev er that this lev el p erceiv es a disk as a linear arra y of blo c ks If

a Unix le system exists on the disk it can b e destro y ed b y acciden tally o v erwriting ino de tables or other
crucial information
Executing Direct SCSI Commands
Normally SCSI devices are con trolled b y a devicet yp esp eci c driv er This device driv er then presen ts the
usual deviceasale abstraction to the programmer With HPUX ho w ev er it is p ossible to explicitly
send SCSI commands to a device through the ioctl system call This system call requires the follo wing
parameters
A le descriptor that w as obtained b y op ening the c haracter sp ecial device with the open system
call
A call iden tication The ioctl system call is v ery generic and can b e used to p erform a v arietyof
functions on c haracter sp ecial les In HPUX the macro SIOC IO iden ties the programmers in ten t
to directly execute a SCSI command
A data structure that con tains the input and output parameters for the request In HPUX the elds
of a structure of t yp e sctl io m ust b e lled in
F or HPUX some of the necessary information on ho w to set up a SCSI io ctl call is con tained in the
scsih header le in the usrincludesys directory Additional information can b e found in the system
man ual pages of diskio scsi and scsi ctl The references describ ed so far only address the question of
ho w to execute a SCSI io ctl call The syn tax and seman tics of the actual SCSI commands are dened in
the SCSI standard do cumen t ANS There are three t yp es of SCSI commands those of length length
and length b ytes The rst b yte of the SCSI command is called the op er ation c o de and uniquely
iden ties the SCSI function The rest of the b ytes are used for parameter passing The standard also
allo ws v endorsp ecic extensions for some of the commands whic h are describ ed in the v endors tec hnical
man uals of the device in question F or example for our exp erimen ts in v olving the Seagate STW
disk driv e w e consulted the tec hnical man uals Seab and Seaa
Bit
Byte
0 PS Reserved Page code (08h)
76 5 43210
1 Page length (0Ah)
2 Reserved WCE MF RCD
3
4
5
6
7
8
9
10
11
Demand read retention priority Write retention priority
(MSB)
(MSB)
(MSB)
(MSB)
(LSB)
(LSB)
(LSB)
(LSB)
Disable pre-fetch transfer length
Minimum pre-fetch
Maximum pre-fetch
Maximum pre-fetch ceiling
Figure Structure of the cac hing page as dened in the SCSI standard for direct access devices
Example Enabling and Disabling the OnDevice Read Cac he
Man y of the new er magnetic disk driv es are equipp ed with cac he memory on the con troller b oard The
SCSI standard ANS sp ecies that it should b e p ossible to enable and disable the cac he function If
an ondevice cac he exists then the related parameters are group ed together in to a blo c k termed a c aching
p age Figure sho ws the structure of the cac he parameter page from the SCSI standard do cumen t ANS
Tw o con trol bits within this page are used directly to enable and disable cac hing
W CE b yte bit Write Cac he Enable This bit enables the write cac hing when set to one
R CD b ytebit Read Cac he Disable This bit disables the read cac hing when set to one
Figures and sho w sample co de fragmen ts in C on ho w to issue direct SCSI commands The function
GetCacheParams in Figure uses the SCSI MODE SENSE command to retriev e the cac hing page The
function SetCacheParams depicted in Figure rst calls GetCacheParams to read the curren t cac he
related v alues and then issues a MODE SELECT command to write them bac k to the device after p ossibly
c hanging the v alues of the W CE and R CD con trol bits

static
int GetCacheP ar ams fd buff
This function issues a MODE SENSE command to the device whose file descriptor is fd It returns the parameter page that de scribes the cache The returned data consists of two parts a byte mode paramete r sense header and
the to byte cache parameter page
These two parts are returned in a modeb uff no b d structure int fd File descript or struct modebu ff no bd buff Caching page The following is adapted from man scsictl memsetbuff BUFFLEN
memsetsct l io sizeofsc tl io Clear reserved fields sctliofla gs SCTLREAD Input data is expected sctliocdb CMDmodes ens e MODE SENSE command sctliocdb x BDB Disable block descriptor sctliocdb x CACHINGPA GE Ask for caching page sctliocdb x Reserved sctliocdb BUFFLEN Allocation length sctliocdb x Control sctliocdb l eng th byte command sctliodat a buff Data buffer location sctliodat a len gth BUFFLEN Maximum transfer length sctliomax m sec s Allow seconds for command if ioctlfd SIOCIO sctlio Request was invalid printfE rr or request was invalidn return
else if sctliocd b sta tu s SGOOD Device is ready return else Unknown state or device is not ready printfE rr or unknown state device is not readyn return
Figure Sample co de fragmen t to retriev e the cac he parameter page of a disk driv e
Exp erimen tal T ec hniques
Quan tifying the Ov erhead of Seek Op erations
T o maximize disk utilization read and write op erations m ust b e carefully sc heduled The more information
the sc heduling algorithm has ab out the seek prole of a disk the tigh ter the disk op erations can b e
sc heduled otherwise w orst case v alues m ust b e assumed and the higher the o v erall utilization of the
system A test program that measures the seek prole of a magnetic disk can therefore pro vide useful

int SetCacheP ar ams fd bWce bRce
This routine can setclea r the WCE Write Cache Enable and the RCD Read Cache Disable bits on a SCSI disk It will return on
success and on error int fd File descripto r
int bWce enable disable write cache int bRce enable disable read cache unsigned char parmlen
struct modebu ff no bd buff
Get the current cache page parameter s if GetCacheP ara ms fd buff parmlen buffmodep ar ms le ngth if parmle n ! BUFFLEN printfE rro r buffer size mismatchn return
else return
Setreset the cache control flags bWce enables cache therefore we must SET WCE if bWce buffcach in gp ag ec on tro lb it s WCE
else buffcach in gp ag ec on tro lb it s "WCE
bRce enables cache therefore we must CLEAR RCD if bRce buffcach in gp ag ec on tro lb it s "RCD
else buffcach in gp ag ec on tro lb it s RCD
Now write the page memsetsct l io sizeofsc tl io Clear reserved fields sctliofla gs x No input data is expected sctliocdb CMDmodes ele ct MODE SELECT command sctliocdb x PF Page format sctliocdb x Reserved sctliocdb x Reserved sctliocdb parmlen Parameter list length sctliocdb x Control sctliocdb l eng th byte command sctliodat a buff Data buffer location sctliodat a len gth parmlen Maximum transfer length sctliomax m sec s Allow seconds for command if ioctlfd SIOCIO sctlio Request was invalid printfE rr or request was invalidn return
else if sctliocd b sta tu s SGOOD Device is ready return else Unknown state or device is not ready printfE rr or unknown state device is not readyn return
Figure Sample co de fragmen t to enabledisable read cac hing of a disk driv e

information for system design
The IO library function call lseek of the C programming language cannot b e used to mo v e the
disk arms directly b ecause it only up dates an in ternal data structure Only the next call to read will
actually execute b oth the seek and the read op eration Ho w ev er timing this call to read will also include
rotational latency dela ys that will bias the measured seek time F ortunately the SCSI Standard ANS sp ecies an explicit seek op eration Through the lo wlev el SCSI in terface describ ed in Section w ew ere
able to force the disks to p erform seek op erations without reading an y data With this command the disk
op eration is nished as so on as the heads reac h the target trac k
The test strategy to measure the seek proles of the magnetic disks in System and System consisted
of a lo op that executed the follo wing four steps during eac h iteration
Seek to logical blo c k address LBA zero
Get the initial time
Seek to LBA X Calculate the total seek time b y subtracting the initial time from the curren t time
The v alue X in step w as v aried from to of the disk capacit y in incremen ts F or ev ery
distinct v alue of X the four steps w ere rep eated times to get an a v erage v alue Figure sho ws the seek
proles of the disks in System and System W e can mak e the follo wing observ ations from the seek proles The seek times are sho wn as a function
of the n um b er of logical blo c ks tra v ersed or the p ercen tage of disk capacit y While this is related to the
n umberofph ysical cylinders w e could not access the ph ysical cylinders directly except for the Seagate
STW disk see the notc h page discussion in Section Since the seek parameter is a n um ber of
logical blo c ks disks that are zoned pro duce dieren t seek proles when seeking in w ard as compared to
out w ard see Figure e This is b ecause a zone with a higher densit ywill con tain few er cylinders for a
giv en n um b er of logical blo c ks when compared to a lowdensit y zone A seek tra v ersing a higher densit y

20 40 60 80 100
Disk Capacity [%]
5
10
15
20
25
30
Seek Time [ms]
20 40 60 80 100
Disk Capacity [%]
5
10
15
20
25
30
Seek Time [ms]
Figure a Quan tum PDS disk on Sys
temF ast SCSI in terface
Figure b HewlettP ac k ard C disk
on System F ast SCSI in terface
20 40 60 80 100
Disk Capacity [%]
5
10
15
20
25
30
Seek Time [ms]
20 40 60 80 100
Disk Capacity [%]
5
10
15
20
25
30
Seek Time [ms]
Figure c HewlettP ac k ard C disk on
System F ast WideSCSI in terface
Figure d Seagate STW disk on Sys
temF ast Wide SCSI in terface
20 40 60 80 100
Disk Capacity [%]
5
10
15
20
25
30
Seek Time [ms]
Outwards Direction
Inwards Direction
Figure e Seeking in w ard vs out w ard with Seagate STW disk
Figure Seek proles for magnetic disk driv es from dieren tv endors All seeks w ere initiated from logical
blo c k address zero

zone will in v olv e a smaller distance reducing the service time of the op eration Seek graphs rep orted b y
man ufacturers are often in terms of cylinders In that case the seek curv es in b oth directions are iden tical
Ho w ev er the n um b er of cylinders p er zone is not alw a ys pro vided b y the disk man ufacturers Section describ es a tec hnique to compute zoning information b y conducting exp erimen ts
Figures bd sho w the seek prole of curren t widely a v ailable GByte disk driv es from Seagate and
HewlettP ac k ard Their maxim um seek time is roughly msec The sligh tly older Quan tum MByte
disk of Figure a executes a fullstrok e seek in ab out msec The seek proles as sho wn in these gures
can b e used as system design parameters to ne tune the realtime sc heduling of disk IOs During our
exp erimen ts b oth test systems w ere in singleuser mo de If the mac hines op erate in normal m ultiuser
mo de system activities suc h as net w ork IOs can cause dela ys due to in terrupt pro cessing or task switc hes
These artifacts mustbetak en in to accoun t when designing a system
Measuring the T ransfer Rate
The dieren t zones of a disk presen t diering data transfer rates to the system W ew an ted to c haracterize
the disk based on its n um b er of zones and the transfer rate storage capacit y and b oundaries of eac h zone
Our exp erimen ts underto ok to measure these disk parameters
The fundamen tal assumption w as that dierences in observ ed transfer rates from the disk w ould rev eal
the desired zoning information In these exp erimen ts w e emplo y ed the rawin terface to the disk to b ypass
HPUX k ernel buering There w as no need to use the lo wlev el SCSI op erations with the exception
of the command to enabledisable the diskcon troller read cac he as describ ed in Section for these
exp erimen ts the standard C library functions read and lseek w ere used
The exp erimen ts w ere conducted on the disks on our t w o platforms b y p erforming an initial lseek
to logical blo c k zero and then sequen tially reading a xed n um ber of b ytes from the disk while timing
eac h read The sequen tial reads w ere con tin ued with no in terv ening lseeks un til the end of the disk

w as reac hed This en tire pro cess w as rep eated times for eac h xed read size The read size w as v aried
bet w een iterations sizes of K K and MBytes w ere used Read sizes w ere selected to b e sucien tly
large so as to negate the eect of the diskcon troller cac he The program stored the start and end times
for eac h read only after all the reads for eac h size iteration w ere completed did the program calculate the
read times W e used this strategy in an eort to minimize the impact of additional pro cessing during the
sequen tial reads W e computed the maxim um minim um and a v erage of the measured times
The transfer rate exp erimen ts w ere p erformed using HPUX in a single user mo de The exp erimen ts
w ere duplicated with the diskcon troller readcac he enabled and disabled so that w e migh t understand the
p ossible eects due to cac hing op erations
200 400 600 800 1000
Disk Capacity [MB]
1.5
2
2.5
3
3.5
4
4.5
5
Transfer Rate [MBps]
200 400 600 800 1000
Disk Capacity [MB]
1.5
2
2.5
3
3.5
4
4.5
5
Transfer Rate [MBps]
Figure a Quan tum PDS disk on Sys
temF ast SCSI in terface
Figure b HewlettP ac k ard C disk
on System F ast SCSI in terface
200 400 600 800 1000
Disk Capacity [MB]
1.5
2
2.5
3
3.5
4
4.5
5
Transfer Rate [MBps]
200 400 600 800 1000
Disk Capacity [MB]
1.5
2
2.5
3
3.5
4
4.5
5
Transfer Rate [MBps]
Figure c HewlettP ac k ard C disk on
System F ast WideSCSI in terface
Figure d Seagate STW disk on Sys
temF ast Wide SCSI in terface
Figure T ransfer rate proles for magnetic disks from dieren tv endors The read cac he w as on in all
cases

Zone Size MB T ransfer Rate MBps Start MB End MB
T able Zone information of Quan tum PDS disk
The transfer rate proles are sho wn in Figure while zone information from our exp erimen ts is
tabulated in T ables The presence of zones manifested as horizon tal segmen ts with diering transfer
rates ma y b e seen in the transfer rate proles The n umberofzonesma y b e determined b y the n um ber of
horizon tal segmen ts in the plot The actual transfer rates themselv es can b e estimated from these plots
but ma y also b e determined directly from the data les
It is of in terest to note that these plots are for a read size of MBytes as this read size pro duces the
cleanest plots for the determination of the n um b er of zones and their sizes Ho w ev er the plots for the
smaller read sizes sho w ed sligh tly higher transfer rates as w ell as an increased scatter in the v alues of the
transfer rates This is b ecause within a giv en zone a large read size will incur appro ximately the same
n um b er of trac k and head switc hes during its retriev al whereas the retriev al of a smaller read blo ckma y
sho w a larger v ariance in these v alues b ecause it ma y sometimes reside en tirely within a giv en trackand
other times on a cylinder b oundary In our exp erimen ts the HP C unexp ectedly pro duced diering transfer rates with cac he on and
o while the transfer rates of the Quan tum PDS and Seagate STW w ere unaected b y cac he
status F rom our exp erimen ts with cac he con trol see Section w e feel that there is some unkno wn to
us in teraction in the HP C b et w een the status of cac he enabledisable and other cac herelated disk
op erations suc h as readaheadprefetc h or zerolatency read that c hanges the observ ed transfer rate
The man ufacturer of the HP C publishes the n um b ers of zones for its device The computed

Zone Size MB T ransfer Rate w T ransfer Rate wo Start MB End MB
Read Cac he MBps Read Cac he MBps
T able Zone information of HewlettP ac k ard C disk
Zone Size MB T ransfer Rate w T ransfer Rate wo Start MB End MB
Read Cac he MBps Read Cac he MBps
T able Zone information of HewlettP ac k ard C F ast Wide disk
Zone Size MB T ransfer Rate MBps Start MB End MB
T able Zone information of Seagate STW F ast Wide disk

n um b er of zones using exp erimen ts matc hed those pro vided b y the man ufacturer with accuracy Our
measured transfer rates for Seagates and HPs transfer rate are not iden tical to those adv ertised b y the
man ufacturers This is b ecause the adv ertised n um b ers corresp ond to ra w data transfer from the disk
medium and include all bits read from the disk surface including the address bits and the ECC bits for
error correction Ho w ev er our measuremen ts are based on the data bits read and include the o v erhead of
SCSI adapter SCSI bus transfer disk con troller disk seeks and head switc hes
F or the Seagate STW disk mo del w ew ere able to v erify our exp erimen tal results ab out zones and
transfer rates with accurate conguration data that w as obtained directly from the devices The STW
disk stores conguration data ab out notches that can b e queried through a SCSI MODE SENSE command
A notc h is the SCSI terminology for a zone This information includes the n um b er of platters the
n um b er of zones and b oth the n um b er of cylinders and the n um b er of logical blo c ks p er zone It is
obtained in a similar manner as the cac he information see Section Using this information w ecan
dev elop analytical mo dels to compute the service time of eachzone seeT able W e can mak e the follo wing observ ations when comparing T ables and W e failed to iden tify t w o zones ie zone of T able w as b eliev ed to b e part of zone in T able and zone of T able w as b eliev ed to b e part of zone in T able b ecause they dier only in one
blo c k p er trac k and therefore their transfer rate is almost the same
W e measured the rest of the zone b oundaries v ery accurately see Figure The measured transfer rate is consisten tly ab out lo w er than calculated This ma y b e due to the
arbitration ie reconnect o v erhead of the SCSI bus proto col
T able enables us to devise a mapping from the logical blo c ks to the ph ysical cylinders This
information could b e used to adjust the seek prole suc h that it is iden tical for in w ard and out w ard
head mo v emen ts
OnDevice Cac he Op eration
In order to supp ort our transfer rate exp erimen ts with the diskcon troller cac he enabled as w ell as disabled
w e underto ok to disco v er the metho ds b y whichw e could send commands to the underlying SCSI hardw are

Zone Cylinders
a
Logical Blo c ks Blks Blks T ransfer Rate
Start End Size Start End Size Cyl T rac k Anal
b
Rat
c
Meas
d
Rat
e
Di
f

a
All zones start on the platter surface of head and end on the platter surface of head b
The analytical transfer rate B l k sT r ack rpm
sec
by tes
Mbyte
c
The ratio b et w een the transfer rate of a giv en zone and the transfer rate of the slowest zone column with fo otnote
b

d
The measured transfer rate see Figure e
The ratio b et w een the transfer rate of a giv en zone and the transfer rate of the slowest zone column with fo otnote
d

f
The p ercen tage dierence b et w een the analytical and the measured transfer rate
Anal T r ansf er Rate M eas T r ansf er Rate
Anal T r ansf er Rate
columns with fo otnotes
b
and
d
T able Zoning information of a Seagate STW disk obtained b y reading the notch p ar ameter p age 200 400 600 800 1000
Disk Capacity [MB]
1.5
2
2.5
3
3.5
4
4.5
5
Transfer Rate [MBps]
Figure Measured transfer rates sup erimp osed o v er accurate zone b oundaries dashed lines for a Seagate
STW disk

Since the Unix op erating system pro vides a deviceindep end en t abstraction of les for accessing devices it
w as necessary for us to circum v en t some standard Unix facilities as detailed in Section The exp erimen ts w ere designed to determine if w e could con trol the op eration of the diskcon troller
cac he b y observing the b eha vior of the transfer rate o v er a small p ortion of the disk W e exp ected to see
am uc h higher transfer rate essen tially that of the SCSI bus when read requests could b e satised using
the diskcon troller cac he instead of the disk medium
50 100 150 200 250
Read Block Size [KB]
0
20
40
60
80
100
120
140
160
Read Time [ms]
Cache enabled
Cache disabled
50 100 150 200 250
Read Block Size [KB]
0
20
40
60
80
100
120
140
160
Read Time [ms]
Cache enabled
Cache disabled
Figure a Quan tum PDS disk on Sys
temF ast SCSI in terface
Figure b HewlettP ac k ard C disk
on System F ast SCSI in terface
50 100 150 200 250
Read Block Size [KB]
0
20
40
60
80
100
120
140
160
Read Time [ms]
Cache enabled
Cache disabled
50 100 150 200 250
Read Block Size [KB]
0
20
40
60
80
100
120
140
160
Read Time [ms]
Cache enabled
Cache disabled
Figure c HewlettP ac k ard C disk on
System F ast WideSCSI in terface
Figure d Seagate STW disk on Sys
temF ast Wide SCSI in terface
Figure Eect of read cac hing for magnetic disks from dieren tv endors
The design of the exp erimentw as to p erform a timed read from the rawc haracter device thereb y
eliminating the in uence of the k ernel buer p o ol Starting at KBytes increasingly larger reads w ere
p erformed in incremen ts of KBytes up to a maxim um of KBytes The reads w ere p erformed from

Disk Measured T ransfer Rate Maxim um SCSI T ransfer Rate
MBytessec MBytessec
Seagate STW HP C HP C Quan tum PDS T able Maxim um Observ ed SCSI T ransfer Rates
blo c k address zero with iterations recorded for eac hreadsize So that w e migh t exp erience disk
con troller read cac he hits when the read cac he w as enabled the exp erimen t rst p erformed a read that
w as not included in the data set times Minima maxima and a v erages w ere calculated for eachset of
ten iterations and plotted in Figure These tests w ere duplicated with the diskcon troller read cac he
enabled and disabled
The in uence of the diskcon troller read cac he is clearly eviden t in the gures sho wing as the segmen t
of the line with a small but nonzero slop e starting from the origin This slop e denes the maxim um rate of
data transfer p er disk These calculated v alues are in T able compared with the corresp onding maxim um
SCSI bus transfer rate
Also eviden t in these gures is the p oin t at whic h the cac he ceases to b e of v alue as indicated b y the
sharp rise in the time required to satisfy a request The v alue of the horizon tal axis at this p oin t is the
size of the read cac he allo cated to our pro cess disk con trollers can segmen t the cac he to pro vide separate
cac he memory for separate pro cesses While w e did p erform tests with the read cac he o as w ell w eha v e
presen ted only those with the cac he on since they sho w ed more in teresting b eha vior
Sev eral other features in Figures ad are of in terest The most ob vious is the o v erall shap e of the
plot consisting of a step function of the time required to read a xed amoun t of data It is curious that
the dierence in time b et w een one plateau and another is equal to the rotational p erio d of the disk While
weha v e no clear explanation of this the corresp ondence is so close that it should b e attributed to some
phenomenon directly related to the rotation of the disk Using these time dierences w e calculated the

Disk RPM estimated RPM Mfg
Seagate STW HP C Quan tum PDS T able Observ ed vs Man ufacturers Sp ecied Disk Rotational Sp eeds
rotational sp eed of the disk These calculated rotational sp eeds compared to the man ufacturers v alues are
in T able Using the Seagate disk w e attempted to disable some of the features w e thoughtma y in uence
the disk b eha vior suc h as prefetc hreadahead Our desired conguration is one in whic h b oth prefetc h
and cac he are disabled ho w ev er our attempts to pro duce this state alw a ys pro duced the same results as
if the cac he w ere enabled
The presence of a handful of p eaks in Figure is not repro ducible from one run to another They
are a consequence of system activit y o ccurring during a timed ev en t and are not an artifact of the disk
hardw are System whic hw as op erated in a standalone mo de exp erienced far few er of these p eaks
than did System whic h remained ph ysically connected to a net w ork Since System is also faster than
System it ma y b e the case that System can pro cess other system activit y with little impact on our
results
Measuremen t of SCSI Scalabilit y
W e conducted exp erimen ts to quan tify the scalabilityc haracteristics of the SCSI bus In these exp eri
men ts w ev aried the n um b er of disks Seagate STW from one to elev en and the size of a read request
from KBytes to MBytes The requests w ere p erformed en tirely within one of t w o zones either the
outermost zone highest transfer rate or the innermost zone lo w est transfer rate
The tests w ere p erformed byha ving a master Unix pro cess fork a c hild pro cess for eac h disk request
The c hild pro cesses p erformed all necessary initialization and then w aited for a sync hronization signal
from the master Up on receipt of the signal eac hc hild issues t w en t y read requests to its assigned disk

one after another After completion of the reads the c hild pro cess a v eraged the measured service times
con v erted them to transfer rates and then wrote them to an output le All tests w ere p erformed with
the diskcon troller read cac he enabled
Disks
T ransfer Rate Innermost Zone MBytessec Outermost Zone T able T otal SCSI bus transfer rate
T able presen ts our results for MByte read requests issued to b oth the innermost and outermost
zones for v arying n um b ers of disks In the innermost zone transfer rate increases linearly up to sev en
disks and then starts to atten out as the bus b ecomes a b ottlenec k W e observ ed similar b eha vior for
read requests to the outermost zone except that the linear region extends to only four disks
0
50
100
150
200
250
0 2 4 6 8 10 12
Standard Deviation of Transfer Times [millisec]
Number of Disks
Outermost Zone
Innermost Zone
Figure Standard deviation of transfer times vs n um b er of disks
Figure depicts the increase in the standard deviation of the read time as the n um b er of disks is
increased This deviation is dep enden t on the n um b er of disks the zone that a blo c k is read from and
the priorit y of the SCSI disk as determined b y its SCSI id In the outermost zone when more than four
sim ultaneous read requests issued the v ariation in the measured service times b ecomes to o large to mo del
accurately Reads from the innermost zones on all disks sho w ed small amoun ts of standard deviation up

to six disks This should come as no surprise b ecause the bandwidth of the bus is xed and the innermost
zone has a lo w er transfer rate than the outermost one
These exp erimen ts demonstrate that mo delling the service time of a SCSI disk is dicult when the
SCSI bus b ecomes saturated
Metho dology to Estimate the Service Time
In order to v erify our metho ds for c haracterizing disk prop erties w ecom bined our results obtained from
the seek test the transfer rate test and the cac he exp erimen ts in to a mo del to predict the service time
for sp ecied read requests Our mo del can estimate the service time of the disk subsystem accurately as
long as the maxim um n um b er of activ e disks do es not exceed four W e emplo y ed this mo del to compare
measured service time to predicted service time Our mo del is designed to nev er underestimate the actual
service time F or the m ultimedia example Figure the time p erio d required to read blo c ks from the
disks w ould nev er exceed the estimates using our mo del
20 40 60 80 100
Disk Capacity [%]
0
5
10
15
20
25
30
Seek Time [ms]
Modeled Seek Profile
Measured Seek Profile
Figure Comparison of measured vs mo deled seek prole for the HewlettP ac k ard C wide disk
driv e

Seek Time
The mo del w as pro duced b y tting a curv e to the seek prole obtained for the in w ard direction The curv e
w as brok en in to t w o parts joined at of disk capacit y The range w as sim ulated with seek time
b eing a square ro ot function of the seek distance while the range from to w as represen ted as seek
time b eing a linear function of seek distance see R W Constan ts w ere c hosen to pro duce the follo wing
seek time equations as w ell as the curv e seen in Figure t ! p
S Capacit y
! S Capacit y
In the ab o v e equations t is the seek time in milliseconds and S is the seek distance expressed in p ercen tage
of disk capacit y Rotational Latency
Added to the mo del w as the calculated v alue for the rotational latency of the disk Ev en though the
rotational latency will often b e less than this v alue onehalf of the rotational latency on the a v erage
c ho osing the full v alue means w e will nev er underestimate this v ariable latency T ransfer Rate
The transfer rates for the observ ed zones on the disk w ere tak en from one of the corresp onding T ables or F or example for the HP C F ast Wide disk driv e w e used column in T able Before
adding the transfer rates to the mo del w e derated them b y Again this derating w as p erformed to
ensure that wew ould nev er underestimate the disk service time
SCSI Scalabilit y
W e used the results from Section to limit the n um b er of disks for whic h our mo del is v alid Based on
the results obtained reading in the outermost highest transfer rate zone w e limit the applicabilit y of our

Read Size Measured Time Predicted Time Error
msec msec p ercen t
K
K
K
K K T able Comparison of measured vs predicted times for starting in Zone and reading in Zone mo del to four disks While up to six disks can b e supp orted if all disks are reading from the innermost
slo w est transfer rate zone our in tended applications m ust use a constantn um b er of disks This limitation
requires that wec ho ose the minim um n um b er of disks that can b e supp orted at all times
Measuremen ts
Since wew an ted to sim ulate the conditions under whic h a disk w ould b e used the follo wing pro cedure w as
emplo y ed to obtain the measured service time
Seek to a sp ecied LBA logical blo c k address on the disk using the lo wlev el SCSI commands This
ph ysically p ositions the disk arms
P erform an lseek to the start LBA of the desired data No disk arm mo v emen t results but the
in ternal le p oin ter is up dated and disk arm mo v emen t will result up on issuing a read command
Obtain the start time
P erform a read for the desired data The disk arms will then seek to the desired lo cation the
platter will rotate un til the desired data is under the disk heads and nally the data will b e read
Compute the elapsed time
This pro cedure w as em b edded within a lo op and p erformed times for eac h set of parameters These
parameters consisted of the initial arm p osition the start p oin t of the read and the n um ber of b ytes to
read
The results of b oth the measured and predicted times as w ell as relativ e errors are presen ted in
T ables and and in Figure F rom Figure w e can see that for up to four disks the service times

Read Size Measured Time Predicted Time Error
msec msec p ercen t
K
K
K
K K T able Comparison of measured vs predicted times for starting in Zone and reading in Zone 5
10
15
20
Zone No.
2
4
6
8
Read Size [MB]
0
10
20
Error [%]
5
10
15
20
Zone No
0
10
20
Figure Relativ e errors ie o v erestimations of mo deled vs measured service times for the Seagate
STW disk driv e
for m ultiple disks are the same as the service times for a single disk As can b e seen from T ables and weha v e a mo del that do es not underestimate the measured service time of a disk y et pro vides a reasonable
appro ximation to the measured service time A more comprehensiv e comparison w as done for the Seagate
STW disk and the results are sho wn in Figure It is eviden t that the estimates in T ables and and Figure are w orse for the smaller read sizes This is b ecause a small read request has a short transfer
time and our estimate is in uenced b y our c hoice of including one full rotational p erio d in our mo del

Conclusions and F uture Researc h Directions
In this study w e used a lo wlev el SCSI programming in terface to iden tify ph ysical attributes of otheshelf
magnetic disk driv es using HPUX based w orkstations This w as p ossible without access to proprietary
information Wew ere able to gather information ab out the seek c haracteristic the zoning and the on
device cac he op eration for sev eral dieren t SCSI disks The exp erimen tal results w ere used to outline
a metho dology to estimate the service time of a disk read request This metho dology can b e emplo y ed
in m ultimedia applications in order to b est utilize magnetic disk driv es while guaran teeing a con tin uous
media displa y Our exp erimen tal tec hniques yield a mo del that predicts the service time of a single disk quite w ell
The main limitation of extending our mo del to m ultiple disks is the uncertain tyin v olv ed in the scalabilit y
and con ten tion in v olv ed using the SCSI bus If w ew ere to extend our mo del past four disks without
underestimating the disk service time w ew ould ha v e had to grossly o v erestimation the disk service time
This w ould w aste the bandwidth of the SCSI disk subsystem
This study further demonstrates that a programmer can ha v e negrained con trol o v er SCSI devices
without ha ving to abandon a con v enien t otheshelf programming en vironmen t Ho w ev er it also sho ws
the limitations of suc h an approac h The m ultiuser m ultitasking nature of Unix in tro duces infrequen t
but hard to con trol dela ys in the disk service time The system designer will ha vetoev aluate these trade
os and w eigh the con v enience of using standard hardw are and soft w are against the stringency of high
disk utilization and lo w hiccup probabilit y of the target application If the highest p ossible n um ber of
sim ultaneous displa ys and absolute hiccupfree retriev al are required then sp ecialpurp ose hardw are and
soft w are m ust b e emplo y ed If on the other hand a lo w er n um b er of displa ys can b e tolerated b y using
the service time estimation tec hniques outlined in this pap er then the use of otheshelf to ols can greatly
reduce the system dev elopmen t eort Moreo v er if there are only a few hardw are and op erating system
dep enden t routines an application p ort to a dieren t platform will b e simplied

Also while the exp erimen ts that w e conducted for this study sho w ed promising results they do not
ha v e the complexit y of a real application The implemen tation of a system that supp orts the displa y
of con tin uous media based on the ideas presen ted in this study will therefore b e another future researc h
direction
References
ANS American National Standard of Accredited Standards Commi tee X Smal l Computer System Interfac e
SCSI ANSI X x Marc h BGMJ S Berson S Ghandeharizadeh R Mun tz and X Ju Staggered Striping in Multimedia Information
Systems In Pr o c e e dings of the A CM SIGMOD International Confer enc e on Management of Data CL HJ Chen and T Little Ph ysical Storage Organizations for TimeDep enden t Multimedia Data In
Pr o c e e dings of the F oundations of Data Or ganization and A lgorithms F ODO Confer enc e Octob er
Ded J Dedek Basics of SCSI Se c ond Edition Ancot Corp oration oST National Institute of Standards and T ec hnology National Storage Industry Consortium In Workshop
on Digital Data Stor ageMarc h P ol VG P olimenis The Design of a File System that Supp orts Multimedia T ec hnical Rep ort TR
ICSI R W C Ruemmler and J Wilk es An In tro duction to Disk Driv e Mo deling IEEE ComputerMarc h Seaa Seagate T ec hnology Inc Disc Drive SCSISCSI Interfac e Pr o duct Manual Volume Version Seab Seagate T ec hnology Inc Hawk LP Family Pr o duct Manual Volume Wil N C Wilhelm A General Mo del for the Performance of Disk Systems Journal of the A CM Jan uary

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 659 (1997)

PDF

USC Computer Science Technical Reports, no. 625 (1996)

PDF

USC Computer Science Technical Reports, no. 627 (1996)

PDF

USC Computer Science Technical Reports, no. 685 (1998)

PDF

USC Computer Science Technical Reports, no. 612 (1995)

PDF

USC Computer Science Technical Reports, no. 618 (1995)

PDF

USC Computer Science Technical Reports, no. 600 (1995)

PDF

USC Computer Science Technical Reports, no. 622 (1995)

PDF

USC Computer Science Technical Reports, no. 598 (1994)

PDF

USC Computer Science Technical Reports, no. 601 (1995)

PDF

USC Computer Science Technical Reports, no. 748 (2001)

PDF

USC Computer Science Technical Reports, no. 623 (1995)

PDF

USC Computer Science Technical Reports, no. 628 (1996)

PDF

USC Computer Science Technical Reports, no. 766 (2002)

PDF

USC Computer Science Technical Reports, no. 666 (1998)

PDF

USC Computer Science Technical Reports, no. 864 (2005)

PDF

USC Computer Science Technical Reports, no. 578 (1994)

PDF

USC Computer Science Technical Reports, no. 653 (1997)

PDF

USC Computer Science Technical Reports, no. 619 (1995)

PDF

USC Computer Science Technical Reports, no. 584 (1994)

Description

Shahram Ghandeharizadeh, James Stone, Roger Zimmermann. "Techniques to quantify SCSI-2 disk subsystem specifications for multimedia." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 610 (1995).

Asset Metadata

Creator Ghandeharizadeh, Shahram (author), Stone, James (author), Zimmermann, Roger (author)

Core Title USC Computer Science Technical Reports, no. 610 (1995)

Alternative Title Techniques to quantify SCSI-2 disk subsystem specifications for multimedia (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 32 pages (extent), technical reports (aat)

Language English

Unique identifier UC16270253

Identifier 95-610 Techniques to Quantify SCSI-2 Disk Subsystem Specifications for Multimedia (filename)

Legacy Identifier usc-cstr-95-610

Format 32 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)