Co esponding au ho : Aguba a Immacula e Chidimma
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion Liscense 4.0.
Enhancing Cache Pe o mance Th ough FIFO eplacemen algo i hm op imiza ion o
high-pe o mance compu ing
Immacula e Chidimma Aguba a *, Ma y O u u Kama and Dickson Apaleokhai Dako
Depa men o So wa e Enginee ing, Facul y o Na u al and Applied Science, Ve i as Uni e si y Abuja, FCT, Abuja, Nige ia.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
Publica ion his o y: Recei ed on 27 June 2025; e ised on 11 Augus 2025; accep ed on 13 Augus 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.27.2.2918
Abs ac
Cache eplacemen algo i hms a e used o op imize he ime aken o he cen al p ocessing Uni (CPU) o p ocess
in o ma ion by s o ing in o ma ion needed by he p ocesso a ha ime and possibly in u u e so ha i he p ocesso
needs ha in o ma ion i can be p o ided immedia ely. The e a e a numbe o echniques (FIFO, LRU, CC, LFU, LRU,
GDSF, MRU, Hyb id) ha a e easily used o o ganize in o ma ion in such a way ha he in o ma ion needed by he CPU
o emain busy and main ain i s speed o p ocessing is eadily a ailable. FIFO is known o i s ease o implemen a ion
and low compu a ional complexi y. The e o e, his pape examines he scena ios in which FIFO ails and p oposes
echniques o enhance cache pe o mance by op imizing FIFO h ough he in eg a ion o machine lea ning o p edic
u u e eques and pe o m in elligen p e e ching o p eload ele an da a in o he cache. The machine lea ning
app oach inc eased he cache hi and inc eased he hi ime.
Keywo ds: Fi s -In-Fi s -Ou ; High Pe o mance Compu ing; Cache; Replacemen Algo i hm; Op imiza ion
1. In oduc ion
Nea he CPU (cen al p ocessing uni ), cache memo y is a iny, as memo y uni ha holds equen ly accessed
in o ma ion and commands. By se ing as a bu e be ween he p ocesso and main memo y also known as Random
Access Memo y (RAM), i speeds up da a e ie al and enhances sys em pe o mance. The cache memo y is o en
cha ac e ized by high speed memo y access, small size and abili y o s o e equen ly used da a. Usually, he cache
memo y is di ided in o h ee le els and each is labelled in ascending o de acco ding o he closeness o he cen al
p ocessing uni (CPU) [1].
When he CPU needs da a, i i s checks he cache memo y and i he da a is ound, i is e ie ed quickly (cache hi )
o he wise, he CPU e ches he da a om he main memo y and s o es a copy in he cache o u u e access.
Since cache memo y g ea ly inc eases sys em e iciency and da a access speeds, i is essen ial o High-Pe o mance
Compu ing (HPC). Wo kloads in HPC sys ems include big da ase s, equen memo y accesses, and in ica e
compu a ions.
A cache eplacemen algo i hm is he p ocess in cha ge o selec ing an i em om he cache o be emo ed and
subs i u ed wi h mo e popula i em. The main aim o he cache eplacemen algo i hms is o maximize he cache hi
a io in o de o imp o e o he pe o mance measu es. They di e in pa ame e s used o selec he i em o be e ic ed
om he cache and he way he pa ame e s a e applied. Some Cache eplacemen pa ame e s include:
•Cache miss: This e e s o an inciden whe e he da a is no ound in he cache.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
907
• Cache hi : This is a si ua ion o incidence whe e he da a is ound.
• Hi ime: Re e s o he ime i akes o access he cache.
• Hi Ra io: Re e s o he numbe o imes he da a is ound in he cache [2].
Some gene ally known cache eplacemen algo i hms include
1.1. Fi s in, i s Ou
Cache memo y is use ul because o he ollowing easons:
• Lowe ing La ency: F equen ly used da a is kep close o he p ocesso because cache memo y may be accessed
a mo e quickly han main memo y (RAM), which lowe s la ency.
• Reducing bo lenecks: Memo y access speeds equen ly es ic CPU pe o mance. This p oblem is lessened
by cache, which o e s quick access o da a and ins uc ions ha a e o en used.
• Inc easing Th oughpu : Pa allel p ocessing is essen ial o HPC wo kloads. Ins ead o wai ing o da a,
p ocesso s can spend mo e ime compu ing when cache is managed e ec i ely.
Cache memo y has an impo an ole o play in high pe o mance compu ing because cache memo y has a ini e amoun
o space, some s o ed da a mus be swapped ou o make oom o new da a when i ills up. To maximize pe o mance,
cache eplacemen algo i hms decide which da a should be emo ed. These me hods a e essen ial o HPC because hey
gua an ee ha high-p io i y o equen ly eques ed da a s ays in cache, minimizing he amoun o expensi e memo y
accesses.
2. Li e a u e e iew
Algo i hmic op imiza ion has been widely applied beyond caching sys ems o imp o e compu a ional e iciency in
a ious domains. Fo ins ance, Aguba a e al. [18] applied Dijks a’s algo i hm in he design and op imiza ion o a bus
booking sys em, signi ican ly educing a el ou e compu a ion imes and imp o ing scheduling e iciency. Al hough,
hei applica ion a ge ed anspo a ion managemen , he p inciple o le e aging op imized algo i hms o minimize
la ency and imp o e se ice deli e y pa allels he objec i es in cache eplacemen op imiza ion, pa icula ly in HPC
en i onmen s. Cache eplacemen algo i hms a e essen ial o managing limi ed cache s o age by de e mining which
da a o e ic when new da a needs o be loaded. Common s a egies include Leas Recen ly Used (LRU), Leas F equen ly
Used (LFU), and Fi s -In-Fi s -Ou (FIFO). FIFO e ic s he oldes da a i s , ope a ing on a simple queue-based p inciple.
While s aigh o wa d and e icien , FIFO doesn' accoun o da a access pa e ns, which can lead o subop imal
pe o mance compa ed o mo e adap i e algo i hms like LRU. Howe e , FIFO's simplici y allows o lowe o e head
and easie implemen a ion, making i sui able o ce ain applica ions whe e access pa e ns a e p edic able o he
o e head o mo e complex algo i hms is p ohibi i e. Recen esea ch has explo ed combining FIFO wi h o he s a egies
o balance simplici y and pe o mance. Fo ins ance, he S3-FIFO algo i hm in eg a es FIFO queues wi h mechanisms o
il e ou in equen ly accessed i ems, achie ing be e scalabili y and h oughpu in web caching scena ios.
In 2018, a pape by [2] compa es he di e en cache eplacemen algo i hms used in ideo se ices such as FIFO, LRU,
LFU(Leas equen ly used), OPT, Chunk-based caching (CC), quali y-based caching (QC) and LRU-2. The esea ch
applies a ziph dis ibu ion o simula e ideo popula i y and e alua es each algo i hm unde di e en cache sizes and
eques a es. The di e en eplacemen algo i hms we e simula ed o de e mine hei pe o mance based on ime, ease
o implemen a ion and o he me ics. Howe e , a s udy by [3] p esen s a pe o mance compa ison simula ion o he
se en cache eplacemen algo i hms on a ious in e ne a ic ex ac ed om he public IRcache da ase . The esul s
o his s udy indica e ha he Hi Ra io (HR) pe o mance is s ongly in luenced by cache size, cacheable and unique
eques s. While a esea ch by [4] shows ha caching as a concep can be used o sa e ene gy and s o e up ene gy o
ene gy consump ion de ices. Again [5]in his esea ch used he PGM echnique o imp o e cache pe o mance by
inc easing cache hi . The me hod inc eased he o e all pe o mance o cache hi by 7% bu could only wo k o cache
hi a io. The able below summa ises he e iew based on he algo i hm used, me hodology, limi a ions o FIFO,
Findings and he me i s i inculca es. Cache eplacemen algo i hms a e essen ial o op imizing pe o mance in
compu ing sys ems. Classical policies such as FIFO, LRU, and LFU ha e been ex ensi ely s udied [1]–[5]. Recen wo ks
ha e also explo ed ad anced FIFO op imiza ions [6]–[8], machine lea ning-enhanced cache policies [9]–[15], and HPC-
speci ic cache s a egies [16], [17].
Yang e al. [6] and Akba i Benga [7] demons a ed ha FIFO, when augmen ed wi h hyb id o p io i y mechanisms,
can achie e pe o mance compa able o o exceeding LRU and LFU. Mahni e al. [8] u he op imized ile-le el cache
placemen in HPC sys ems using a mul i-c i e ia app oach, boos ing hi a es and I/O e iciency.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
908
Machine lea ning-based app oaches ha e signi ican ly ad anced cache managemen . Se humu ugan e al. [9], Shi e al.
[10], Vie i e al. [11], Rod iguez e al. [12], Liu e al. [13], Choi and Pa k [14], and Zhou e al. [15] applied p edic i e
models, deep lea ning, and ein o cemen lea ning o an icipa e access pa e ns, he eby enhancing hi a ios and
adap abili y. Again, o HPC wo kloads, Jame e al. [16] and Wu e al. [17] con i med ha ailo ed eplacemen
s a egies can yield subs an ial pe o mance gains in la ge-scale, pe o mance-sensi i e en i onmen s.
Table 1 Summa y o e iew
Au ho (yea
o
publica ion)
Algo i hm
Me hodology
Limi a ions
o FIFO
Findings
Me i s o FIFO
iden i ied.
Zul a e al
(2023)
FIFO, LFU,
LRU, GDSF,
GDS, SIZE,
LFUDA
Pe o med a
pe o mance
compa ison o he
di e en cache
eplacemen
algo i hms on
a ious in e ne
a ic.
FIFO
s uggles
unde high
access
anomalies.
FIFO had lowe hi a io
compa ed o GDSF and
LRU.
FIFO has low
compu a ional
complexi y.
S allings
(2019)
FIFO,
Random,
LFU
Has highe
cache miss
FIFO is p edic able bu
ine icien compa ed o
LRU
Has low
p ocessing
o e head
Osman &
Osman
(2018)
FIFO, LRU,
LFU, LRU-2,
QC, CC, OPT
A simula ion-based
compa ison o
cache eplacemen
algo i hms o
ideo se ices
I is mainly a
simula ion-
based
esea ch
The esea ch ound ha
he CC algo i hm achie es
he la ges hi a io and
pe o ms well e en unde
small cache sizes.
Howe e , he FIFO
algo i hm has he
smalles hi a io among
all algo i hms.
Easy o
implemen
Ali (2016)
FIFO, LRU,
MRU, LIFO,
Hyb id
De eloped an
algo i hm o
inc ease cache hi
h ough educ ion
and main enance o
o e head
FIFO does no
conside
access
equency
FIFO unde pe o ms in
dynamic access pa e ns
FIFO is simple
and incu s low
o e head
Osman
(2016)
No
applicable
Ene gy e iciency o
caching in op ical
ne wo ks
I lacks
con en
popula i y
awa eness
FIFO can educe ne wo k
ene gy consump ion bu
no op imized o la ge
scale
I was no based
on any
algo i hm
Hence, om he li e a u e e iew summa y he ollowing indings can be deduced abou FIFO:
• FIFO is an easy algo i hm o implemen and akes less sys em esou ces and logic o execu e.
• FIFO equi es simple modi ica ions and upg ade o ou pe o m o he algo i hms since i is easy o implemen .
Al hough cache is a e y e icien algo i hm and pe o ms well in mos scena io, i pe o ms poo ly in scena ios whe e
da a access is unp edic able and wo kloads a e dynamic. Some easons why cache may no wo k well includes:
F equen Reuse o Olde Da a (Cache Pollu ion Issue): In da abase cache que ying, olde da a is been used mo e han
he jus ecen ly queued da a, hence FIFO does no apply in his ype o scena io.
Looping o Cyclic Access Pa e ns: FIFO pe o ms poo ly in scena ios ha in ol es a cyclic o looping s uc u e. This is
because FIFO elimina ed he olde cache i ems.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
909
The me hodology is o mula ed in such a way o ensu e ha he ad an ages/bene i s o FIFO is duly ha nessed and
limi a ions i possesses a e managed wi h he imp o ed o hyb id app oaches ha ensu es he bes esul s in cache
pe o mance.
Mul i-le el FIFO (ML-FIFO): This me hodology p oposes o c ea e a mul i-le el o ie ed FIFO queues ha will handle
di e en ypes o da a e icien ly. This includes:
• Sho - e m FIFO cache: This will handle sho ecen eques
• Long- e m FIFO cache: This will handle equen ly accessed da a.
Hence, hese me hodologies ensu e ha when he cache is ull, only he leas impo an da a is e ic ed om he cache.
Dynamic Th eshold FIFO (DT-FIFO): This solu ion wo ks on he p emise ha i access equency inc eases, FIFO
beha es like LRU. Howe e , wo kload is andom, FIFO main ains i s basic o m o i s in, i s ou .
Hyb id Model FIFO (HM-FIFO): This me hodology sugges s ha FIFO be combined wi h o he eplacemen algo i hms
o ge op imal esul . The ollowing combina ions a e conside ed o gi e op imal esul .
FIFO wi h LFU: This me hod will combine he simplici y o FIFO and he high hi a io o LFU; he e o e, his will wo k
in a way ha enables equen ly accessed i ems o s ay longe in he cache while a ely accessed i ems a e eplaced i s .
FIFO wi h LRU: This will combine he abili y o p io i ize LRU abili y o e ain ecen ly accessed da a educing cache
misses. I wo ks by se ing a h eshold whe e FIFO is used ini ially bu is accessed equen ly wi hin a window. I is
mo ed o an LRU-managed sub-cache.
Adap i e FIFO wi h aging mechanism: This in ol es he implemen a ion s a egy ha delays he e ic ion o equen ly
accessed i ems. The me hodology wo ks by assigning weigh s o he di e en cache i ems in he queue. The weigh s will
be assigned in ascending o de om he mos equen ly accessed. As he weigh inc eases, he leas equen ly accessed
i ems inc ease in he alue o hei assigned weigh s, hence he i em in he queue wi h he highes weigh is always he
i em o be e ic ed nex on he FIFO queue. Enabling FIFO check usage pa e ns and ins ead o e ic ing blindly, i delays
he e ic ion o equen ly accessed i ems.
2.1. In elligen P e e ching & Machine Lea ning In eg a ion
P edic i e P e e ching: This basically in ol es he use o his o ical access pa e ns and p edic he u u e o how da a
will possibly be accessed. This includes he p edic ion o p e-load da a based on pas eco ds. The me hod in eg a es
machine lea ning models such as ma ko chains and neu al ne wo ks o an icipa e he cache en ies ha will be needed.
This is pa icula ly use ul in ideo s eaming se ices and cloud compu ing wo kloads.
Rein o cemen Lea ning-Based Cache Replacemen : This imp o emen me hod uses A i icial In elligence- d i en (AI-
d i en) decision making om pas pe o mance o dynamically adjus cache policies. The basic applica ion a eas a e in
da a cen e s and cloud s o age. This me hodology adap s in eal- ime o changing wo kloads.
3. Me hodology
The p oposed Machine Lea ning-Enhanced FIFO (ML-FIFO) amewo k builds upon he simplici y o he Fi s -In-Fi s -
Ou (FIFO) cache eplacemen policy while add essing i s key limi a ion: lack o adap abili y o a ying access pa e ns
in high-pe o mance compu ing (HPC) en i onmen s. By in eg a ing p edic i e analy ics, he sys em es ima es he
likelihood o u u e access o each cached i em and eo de s e ic ion p io i ies acco dingly. The me hodology ollows
he s uc u ed wo k low ou line below.
3.1. Da a Collec ion
To model and e alua e he p oposed ML-FIFO app oach, his o ical access logs a e ob ained om con en Deli e y
Ne wo k (CDN) wo kloads and HPC applica ion aces. Da a aces om public eposi o ies like IRCache and simula ed
HPC job schedule s will be sou ced. Each log en y includes he imes amps o access, objec iden i ie , access equency,
and cache hi /miss s a us. The da ase encompasses di e en a ic scena ios such as sequen ial s eaming, andom
access, and bus y eques pa e ns o e lec ealis ic HPC ope a ional condi ions. Also, logs a e segmen ed in o ixed-
leng h obse a ion windows o cap u e empo al locali y while main aining manageable compu a ional equi emen s.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
910
This da ase o ms he baseline o bo h model aining and pe o mance benchma king.
3.2. Fea u e Enginee ing
Raw access logs a e p ocessed o ex ac p edic i e ea u es ha cap u e bo h sho - e m and long- e m access
beha io . These ea u es include:
• Recency: Time elapsed since he las access o each cached i em.
• F equency: Numbe o accessed wi hin he cu en obse a ion window.
• In e -a i al Time: A e age du a ion be ween successi e accesses.
• Access T end: Di e ence be ween sho - e m and long- e m equency, indica ing inc easing o dec easing
popula i y.
• Tempo al Pa e ns: Time-o -day and wo kload-phase indica o s o accoun o cyclical usage pa e ns in HPC.
All Fea u es a e no malized and encoded o compa ibili y wi h machine lea ning models. Missing alues a e impu ed
using olling a e ages o a oid bias.
3.3. Model T aining
Two complemen a y models a e ained o le e age he ex ac ed ea u es:
• Long Sho -Te m Memo y (LSTM) Ne wo k: Cap u es sequen ial dependencies and long- ange empo al
co ela ions in access pa e ns. T ained using a sliding window sequence app oach, he LSTM p edic s he
p obabili y o a u u e access wi hin a de ined ho izon (e.g., nex n accesses).
• Random Fo es Classi ie : Handles non-linea ela ionships be ween enginee ed ea u es and pe o ms bina y
classi ica ion in o ‘ e ain’ o ‘e ic ’. Fea u e impo ance ankings om he RF model a e also used o alida e
he signi icance o enginee ed ea u es.
3.4. T aining p ocess
Da a is spli in o 80% aining and 20% es ing se s. Hype pa ame e s (e.g., LSTM hidden uni s, RF ee dep h) a e uned
using c oss- alida ion and he models a e e alua ed using p ecision, ecall, and F1-sco e o ensu e balanced p edic i e
pe o mance.
3.5. In eg a ion wi h FIFO
The in eg a ion chosen p ese es FIFO’s O(1) ime complexi y o mos ope a ions while selec i ely al e ing e ic ion
o de o e ain aluable i ems. The s anda d FIFO queue is enhanced wi h a p edic i e decision laye :
• When he cache is ull, he i em a he on o he FIFO queue is examined.
• The ML p edic ion module es ima es i s u u e access p obabili y.
• I he p edic ed p obabili y exceeds a de ined h eshold (e.g., 0.7), he i em is mo ed o he ea o he queue,
delaying i s e ic ion.
• I he p obabili y is below he h eshold, he i em is e ic ed ollowing s anda d FIFO logic.
3.6. E alua ion
ML-FIFO is benchma ked agains s anda d FIFO, Leas Recen ly Used (LRU), and Leas F equen ly Used (LFU) policies.
E alua ion includes s a is ical signi icance es ing o con i m pe o mance imp o emen s. Pe o mance is e alua ed in
a con olled simula ion en i onmen eplica ing HPC wo kload. The ollowing me ics a e measu ed:
• Hi Ra io: Pe cen age o eques s se ed di ec ly om he cache.
• Miss Ra io: Complemen o he hi a io, indic ing cache ine iciency.
• Execu ion Time: Time aken o cache ope a ions, including any addi ional o e head om ML in e ence.
• Th oughpu : Numbe o eques s p ocessed pe second.
3.7. P oposed Me hodology
The esea ch p oposes a machine-lea ning enhanced FIFO o p edic i e p e e ching. The me hod uses machine
lea ning models (long sho - e m memo y (LSTM), andom o es ) o p edic which cached i em will be accessed nex
and FIFO is adjus acco dingly wi h he i ems o be accessed nex , o ganized a he back o he queue. Hence he
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
911
o ganiza ion allows he FIFO algo i hm e ic i ems ha a e no likely o be accessed a he on o he queue whe e can
easily be e ic ed by he algo i hm.
How i wo ks
• S ep 1: The LSTM model is ained on access pa e n using da a om he p e ious o pas access pa e n
gene a ed.
• S ep 2: Be o e an i em is e ic ed, he LSTM model p edic s he u u e access p obabili y o he i em
• S ep 3: Gi en he esul o he p edic ion in s ep 2 abo e, an i em wi h high u u e access p obabili y ge s
e ained i.e. ge s a e en ion boos ins ead o an e ic ion
This will in elligen ly p e en cache misses by e aining impo an i ems longe in he cache. This will be implemen ed
using da a om con en deli e y ne wo k (CDN) whe e i al con en mus s ay in he cache longe . The p oposed
amewo k balances FIFO’s scalabili y and simplici y wi h he p edic i e in elligence o machine lea ning, enabling
adap abili y in highly dynamic HPC en i onmen s wi hou imposing excessi e compu a ional o e head.
4. Discussions
The s udy in oduced a se ies o enhancemen s o he adi ional FIFO cache eplacemen algo i hm, each aimed a
add essing speci ic limi a ions while e aining i s inhe en simplici y and low compu a ional cos . The p oposed
modi ica ions which is hyb id FIFO, Adap i e FIFO, and ML-Based FIFO demons a e how inc emen al imp o emen s
can signi ican ly impac cache pe o mance ac oss di e se applica ion domains. The able below shows he
enhancemen ha we e made o he FIFO eplacemen algo i hm.
Table 1 Enhancemen ha we e made o he FIFO eplacemen algo i hm
P oposed
Modi ica ion
Key change
Bene i
Applica ion a ea
Hyb id FIFO
Keeps T acks o he equency o use o an
i em be o e e ic ion
Reduc ion in he numbe o
unnecessa y emo als
Da abase caching
Adap i e FIFO
Dynamically adjus he e en ion o i ems
by FIFO, so as o a oid o s a ic
o ganiza ion o
FIFO o i s come, i s se e.
This leads o adap a ion o
wo kload changes.
Used in cloud
compu ing
ML-Based FIFO
P edic ion o u u e access o i ems in he
Cache
Gene ally, enhances cache
e iciency
Used in CDNs and AI-
d i en caching
4.1. Hyb id FIFO
Inco po a es equency acking alongside he FIFO queue s uc u e. I main ains an access coun o each cached i em,
he algo i hm can p e en p ema u e e ic ion o equen ly accessed da a. This hyb id app oach educes unnecessa y
emo als and imp o es da a locali y, making i pa icula ly bene icial o da abase caching, whe e epe i i e que ies o
he same da a segmen s a e common. Compa ed o s anda d FIFO, his me hod in oduces minimal addi ional me ada a
bu yields no iceable gains in hi a io.
4.2. Adap i e FIFO
Add esses he s a ic na u e o con en ional FIFO by dynamically adjus ing i em e en ion imes based on wo kload
cha ac e is ics. Ins ead o e ic ion s ic ly on a i s -in- i s -ou basis, he algo i hm adap s i s e ic ion decisions
acco ding o changing access pa e ns. This adap abili y is c i ical in cloud compu ing en i onmen s, whe e wo kloads
can shi unp edic ably due o mul i- enan esou ce alloca ion and elas ic scaling. The obse ed imp o emen in hi
a io o e baseline FIFO con i ms ha esponsi eness o wo kload a iabili y can signi ican ly enhance o e all cache
e iciency.
4.3. ML-Based FIFO
Rep esen s he mos ad anced enhancemen , employing p edic i e modeling o o ecas u u e access p obabili ies.
Ha nessing his o ical access pa e ns, he machine lea ning module eo de s he e ic ion queue, p io i izing he
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
912
e en ion o i ems wi h high likelihood o euse. This app oach is especially well-sui ed o con en deli e y ne wo ks
(CDNs) and AI-d i en caching sys ems, whe e p edic i e in elligence can mi iga e cache misses o ending o high-
demand con en . While his me hod in oduces mode a e compu a ional o e head due o model in e ence, he po en ial
o subs an ial gains in hi a io jus i ies he ade-o in many high-pe o mance applica ions
Table 2 Benchma k Resul s om exis ing algo i hms based on pa ame e s such as hi a e, miss a e and execu ion ime
Algo i hm
Hi Ra e
Miss Ra e
Execu ion Time (s)
S anda d FIFO
40.1%
59.9%
0.0023
Adap i e FIFO
55.6% (Be e )
44.4%
0.0032
Hyb id FIFO-LFU
64.3% (Much Be e )
35.7%
0.0041
LRU (Baseline)
71.2% (Bes Pe o mance)
28.8%
0.0055
LFU (Baseline)
69.4%
30.6%
0.0060
4.4. Pe o mance Analysis
The benchma k esul s (Table 3) highligh he compa a i e pe o mance o he enhanced FIFO app oaches agains
es ablished algo i hms such as LRU and LFU. S anda d FIFO eco ded a hi a e o 40.1%, signi ican ly lowe han LRU’s
71.2%, con i ming he limi a ions o a pu ely sequen ial e ic ion policy. The Adap i e FIFO imp o ed he hi a e o
55.6%, indica ing ha esponsi eness o wo kload dynamics can u he inc eased he hi a e o 64.3%, app oaching
LFU’s 69.4% while main aining lowe execu ion ime. The able below shows he expec ed pe o mance imp o emen s.
Table 3 Expec ed Pe o mance Imp o emen s
Algo i hm
Hi Ra e
Miss Ra e
Execu ion Time
FIFO (Baseline)
40-50%
50-60%
Fas
Adap i e FIFO
55-65%
35-45%
Mode a e
Hyb id FIFO-LFU
60-70%
30-40%
Mode a e
LRU (Baseline)
70-80%
20-30%
High
ML-Enhanced FIFO (P oposed)
75-85%
15-25%
Mode a e (Wi h ML O e head)
The p oposed ML-Enhanced FIFO is p ojec ed o achie e hi a es be ween 75% and 85%, ou pe o ming bo h LRU and
LFU in p edic i e scena ios. Al hough i s execu ion ime is mode a e due o he machine lea ning o e head, i emains
compe i i e o HPC and CDN wo kloads whe e p edic ion accu acy ansla es di ec ly in o pe o mance and cos
sa ings.
5. Conclusion
These indings sugges ha FIFO, o en ega ded as a simplis ic baseline, can be ans o med in o a high-pe o mance
cache eplacemen policy h ough a ge ed enhancemen s. The esul s also emphasis he impo ance o ligh ning
algo i hmic imp o emen s wi h speci ic applica ion domains. Hyb id FIFO o da abases, Adap i e FIFO o cloud
pla o ms, and ML-Enhanced FIFO o in elligen caching in dis ibu ed sys ems. The ade-o be ween compu a ional
o e head and p edic i e accu acy should guide implemen a ion choices, especially in la ency-sensi i e en i onmen s.
Compliance wi h e hical s anda ds
Disclosu e o con lic o in e es
No con lic o in e es o be disclosed.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 906-913
913
Re e ences
[1] W. S allings, Compu e O ganiza ion and A chi ec u e: Designing o Pe o mance, Pea son, 2020.
[2] N. I. O. A eej M. Osman, "A Compa ison o Cache Replacemen Algo i hms o Video Se ices," In e na ional
Jou nal o Compu e Science and In o ma ion Technology, ol. 10, no. 2, pp. 95-111, 2018.
[3] A. E. P. A. F. W. A. Mulki Indana Zul a, "Pe o mance compa ison o cache eplacemen algo i hms on a ious
in e ne a ic," Ju nal In o el, ol. 15, no. 1, pp. 1-7, 2023.
[4] N. I. Osman, "Will ideo caching emain ene gy e icien in u u e co e op ical ne wo ks?," digi al communica ions
and ne wo ks, ol. 3, pp. 39-46, 2017.
[5] A. Sa wan, Cache eplacemen algo i hm, Peshawa , 2016.
[6] J. Yang, Z. Qiu, Y. Zhang, Y. Yue, and K. V. Rashmi, “FIFO can be be e han LRU: The powe o lazy p omo ion and
quick demo ion,” in P oc. Ho OS ’23, 2023.
[7] D. Akba i Benga , “P io i y Cache Objec Replacemen by Using LRU, LFU and FIFO algo i hms o Imp o e Cache
Memo y Hi Ra io,” T ans. So Compu ., ol. 1, no. 1, 2025.
[8] H. Mahni, S. Rubini, S. Gougeaud, P. Deniel, and J. Boukhobza, “Mul ic i e ia File-Le el Placemen Policy o HPC
S o age,” in P oc. SAC ’25, 2025.
[9] S. Se humu ugan, J. Yin, and J. Sa o i, “Designing a cos -e ec i e cache eplacemen policy using machine
lea ning,” in P oc. HPCA, 2021, pp. 291–303.
[10] Z. Shi, X. Huang, A. Jain, and C. Lin, “Applying deep lea ning o he cache eplacemen p oblem,” in P oc. MICRO
’52, 2019, pp. 413–425.
[11] G. Vie i, L. V. Rod iguez, S. Lyons, e al., “D i ing cache eplacemen wi h ML-based LeCaR,” in P oc. Ho S o age,
2018.
[12] L. V. Rod iguez, F. Yusu , S. Lyons, e al., “Lea ning cache eplacemen wi h CACHEUS,” in P oc. FAST ’21, 2021,
pp. 341–354.
[13] E. Liu, M. Hashemi, K. Swe sky, P. Rangana han, and J. Ahn, “An imi a ion lea ning app oach o cache
eplacemen ,” in P oc. ICML, 2020, pp. 6237–6247.
[14] H. Choi and S. Pa k, “Lea ning u u e e e ence pa e ns o e icien cache eplacemen decisions,” IEEE Access,
ol. 10, pp. 25922–25934, 2022.
[15] Y. Zhou, F. Wang, Z. Shi, and D. Feng, “An end- o-end au oma ic cache eplacemen policy using deep
ein o cemen lea ning,” in P oc. ICAPS ’22, 2022, pp. 537–545.
[16] A. V. Jame , L. Al a ez, D. A. Jiménez, and M. Casas, “Cha ac e izing he impac o las -le el cache eplacemen
policies on big-da a wo kloads,” in P oc. IISWC, 2020, pp. 134–147.
[17] H. Wu, Y. Luo, and C. Li, “Op imiza ion o hea -based cache eplacemen in edge compu ing sys em,” J.
Supe compu ., ol. 77, no. 3, pp. 2268–2301, 2021.
[18] I. C. Aguba a, M. Kha a , O. C. Onyedeke, M. Ezema, and C. N. Okwueze, “Design and Op imiza ion o Bus Booking
Sys em using Dijks a’s Algo i hm,” In e na ional Jou nal o Science and Business, ol. 4, no. 12, pp. 21–37, No .
2020, doi: 10.5281/zenodo.4232573.
