Computers & Electronics

Data networks are lightly utilized, and will stay that way

Data netwrks are lightly utilized, and will stay that way Andrew Odlyzk AT&T Labs - Research Octber 7, Abstract. The ppular press ften extlls packet netwrks as much mre efficient
of 19
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Data netwrks are lightly utilized, and will stay that way Andrew Odlyzk AT&T Labs - Research Octber 7, Abstract. The ppular press ften extlls packet netwrks as much mre efficient than switched vice netwrks in utilizing transmissin lines. This impressin is reinfrced by the delays experienced n the Internet and the famus graphs fr traffic patterns thrugh the majr exchange pints n the Internet, which suggest that netwrks are running at full capacity. This paper shws the ppular impressin is incrrect; data netwrks are very lightly utilized cmpared t the telephne netwrk. Even the backbnes f the Internet are run at lwer fractins (10% t 15%) f their capacity than the switched vice netwrk (which perates at ver 30% f capacity n average). Private line netwrks are utilized far less intensively (at 3% t 5%). Further, this situatin is likely t persist. The lw utilizatin f data netwrks cmpared t vice phne netwrks is nt a symptm f waste. It cmes frm different patterns f use, lumpy capacity f transmissin facilities, and the high grwth rate f the industry. 1. Intrductin Annuncements f new packet netwrks ften lead t news stries claiming IP (Internet Prtcl) netwrks are faster and less expensive than traditinal circuit-switched netwrks (cf. [Keller]). Usually n explanatin is ffered fr this claimed advantage f packet transmissin. Mre technical presentatins explain that ld-style phne netwrks reserve tw circuits (ne in each directin) fr a phne call, even thugh almst all the time nly ne persn is speaking, and that there are frequent pauses during cnversatins when nthing is being transmitted. In cntrast, packet netwrks transmit data nly when there is smething t send, and thus it is plausible that they wuld use transmissin capacity mre efficiently. Vint Cerf, ne f the fathers f the Internet, made the fllwing cmparisn f packet versus circuit switching (in the Telecm Italia presentatin at [Cerf]): Circuit (telephny) like reserving bicycle lanes frm LA t NY! Packet (Internet) like sharing f the highway amng high speed cars. That is an appealing analgy. Hwever, it cnceals a much mre cmplicated picture. It appears that tday mst cmpanies are paying mre fr large file transfers ver their private IP-based netwrks than they wuld if they used mdems ver the public switched vice netwrk. This is nt an argument fr circuit-switched netwrks ver packet nes, since there are ther cmpelling arguments in favr f IP netwrks (see the cmpanin papers [Odlyzk1, Odlyzk2]). Hwever, it des suggest the need fr a mre careful investigatin f just hw data netwrks are used. This paper studies average utilizatin levels f transmissin lines in data netwrks, where the averages are ver a full week. Surprisingly, althugh there is a huge literature n netwrks, such averages appear t have been little studied, althugh they are critical t understanding the ecnmics f data netwrks. One minr reasn fr cncentrating n transmissin is that it is the easiest t measure, since switching r ruting capacity is ntriusly hard t quantify. A much mre imprtant reasn is that transmissin is the mst expensive part in a data netwrk. (We cncentrate n lng distance transprt nly, and s d nt take int accunt lcal netwrking csts, such as thse f mdems fr residential custmers f ISPs, which are the bulk f the ttal cst f Internet services such as America Online.) Typical crprate inter-lan netwrks appear t spend arund 45% f their perating expenses n transmissin, 20% n equipment (depreciatin and maintenance) and 35% n peple. One reginal ISP reprts spending 55% f perating funds n transmissin and 15% n equipment. Similar estimates that shw the dminant rle f transmissin csts can be fund in the cst mdel fr ISPs develped by Leida [Leida]. If data netwrks are intensively utilized, then we shuld find transmissin lines run at high fractins f their capacity. That is certainly a widespread view. The impressin that packet netwrks have high utilizatin levels f transmissin and switching facilities is reinfrced by the delays bserved n the Internet (the Wrld Wide Wait ) and the widely publicized data n usage patterns. Figure 1 (based n Fig f [Ash]) shws the traffic n the U.S. switched vice netwrks ver a tw-day perid. It is peaked, as flks in Peria d nt like t call their friends r business partners in Pughkeepsie at 3 am. Thus there are lng perids when that netwrk is largely idle. During the tw days shwn in Fig. 1, the average traffic was abut 40% f the peak. On the ther hand, Fig. 2 shws traffic thrugh the PacBell NAP (Netwrk Access Pint), a majr exchange pint n the Internet, during Octber 26 and 27, (Additinal data fr this NAP, as well as ther exchanges, is available thrugh links prvided at [CAIDA, NLANR].) This NAP was running full blast almst arund the clck. As a fractin f the peak rate bserved during thse tw days, the average thrughput was 84% n Mnday and 80% n Sunday. The perceptin f intensive use even f crprate netwrks is reflected in frequently heard cmments abut 70% utilizatin levels f private lines. These cmments are ften made withut qualificatin, as if they reflected lng-term averages. Mre experienced peple make mre precise statements. 2 Fr example, Fred Baker f Cisc reprts (private cmmunicatin) that crprate custmers cmmnly claim their inter-site WAN links are used at 70% f capacity during peak perids. Brett Leida [Leida] has a mdel fr the lad n a typical T1 line frm a crprate custmer t the Internet which has the peak perid lad at 70% fr several hurs each business day, and average lad f 34%. Leida btained his infrmatin frm members f the MIT Internet Telephny Cnsrtium, which includes many established cmmunicatins industry players. This paper presents extensive evidence that average utilizatin levels are far lwer than generally suppsed. While the lng distance circuit switched vice netwrk has average utilizatin f abut 33%, the Internet backbne links appear t have average utilizatins clser t 10% t 15%, and crprate lng-haul links (which is where the bulk f data transprt capacity is) have utilizatins in the 3% t 5% range. A better analgy than Vint Cerf's might be: Circuit (telephny) like a lane frm LA t NY that is full f well-behaved bicyclists. Packet (Internet) like sharing f the highway amng high speed cars, but with frequent cnstructin deturs. Packet (crprate Intranet) like sharing f a 100-lane highway amng a few high speed cars. At first sight it seems that it shuld be simple t determine average utilizatin levels. That is nt s, thugh, since, fr privacy reasns, carriers such as AT&T d nt mnitr hw the private lines they lease t custmers are used. Individual custmers in many cases d nt measure their wn usage. When they d measure it, they ften d nt btain average utilizatin levels. Even thse statistics that are cllected are usually regarded as cnfidential. Thus it is hard t btain slid estimates, and it is necessary t resrt t limited sampling and circumstantial evidence. The crprate managers wh reprt 70% utilizatin levels are crrect. Their netwrks d generate such figures, but they are usually misinterpreted. Given the way statistics are cllected in many systems, the 70% figure may nt even refer t the busy hur ver a week, but the busiest 5 minutes ver a perid f mnths. Further, it typically applies t nly a few links in a system. Of the varius peple that I have talked t, the nes wh accepted my claims f lw utilizatin levels mst readily were designers f private line netwrks. They are nt used t cnsidering utilizatin rates averaged ver a full week. Hwever, nce I explained t them that this is what I was after, they 3 typically did a quick mental calculatin and said Of curse, this is bvius because f [factrs that will be discussed in Sectin 8 f this paper]. Hwever, such lng-term averages are irrelevant. Lw average utilizatin levels are indeed irrelevant t designers f private line netwrks. These designers have t prvide levels f service specified by their custmers at minimal cst, and lng-run averages d nt matter t them. Hwever, as is shwn in Sectin 9 belw and in the cmpanin papers [Odlyzk1, Odlyzk2], average utilizatin rates are imprtant fr understanding such imprtant questins as the prfitability f the ISP business, the prspects fr packet telephny, and general evlutin f data netwrks, in particular prspects fr Quality f Service. As a sample f the kinds f arguments that can be based n the data in this paper, cnsider Fig. 4, which shws the usage prfile fr a crprate 128 Kbps dedicated cnnectin t the Internet. This business uses the Internet bth fr general cnnectivity, and als t transmit data between different lcatins. The average utilizatin is abut 3%, fairly typical fr such links. This business culd clearly receive all its data n a 56 Kbps link at a cst f suffering delays f at mst minutes, and pssibly nly secnds, in its cmmunicatins. If were all that was being transmitted, that wuld surely be acceptable, and a 56 Kbps link is all that wuld be in place. That this business pays fr a 128 Kbps cnnectin shws that it values the ability t ccasinally send r receive data at high rates. The high speed bursts are extremely infrequent, thugh, and seldm d several cllide t saturate the link. Therefre Quality f Service measures wuld nt be f much help. Further, even if 90% f the traffic n this link were frivlus persnal usage (stck qutes, cartns, and s n), banning it wuld nt prvide significantly better perfrmance fr the high pririty applicatins that justify the cst f the link. When the high pririty traffic starts up, it almst always gets the full bandwidth f the link in any case. Nte that these arguments wuld nt apply if the link were rutinely used at 70% f capacity during business hurs, as is cmmnly believed. In heavy utilizatin cnditins, either Quality f Service measures r banning nn-essential traffic wuld prvide better service fr the missin-critical applicatins. That average utilizatins n data netwrks are lw shws what kind f cnnectins are desired by custmers, and hw highly they are valued. In particular, the lw utilizatin rates d thrw serius dubt n the advisability f many Quality f Service appraches [Odlyzk1, Odlyzk2], which appear t be mtivated by the assumptin that netwrks are heavily cngested. Lw utilizatin rates lead t great pprtunities fr higher quality r less expensive service frm aggregatin f traffic. If tw business custmers have 128 Kbps lines that are used at 70% f capacity during the peak business hurs, relatively little can be gained by cmbining their traffic streams. One wuld still need 256 Kbps f capacity. On the ther hand, if they bth behave like the business f Fig. 4, 4 aggregating their traffic n a 192 Kbps circuit wuld give each ne the perceptin f having a dedicated 192 Kbps link. On larger scales, with mre custmers invlved, the benefits are much greater, and they underlie the ecnmics f public netwrks. Sectin 8 presents quantitative analyses f the reasns fr lw average utilizatin rates f data netwrks, and argues that such rates will persist. The cmpanin paper [Odlyzk1] suggests sme ways t increase thse utilizatin rates t sme extent. Hwever, it is unlikely that data netwrk utilizatin rates will ever apprach thse f the switched vice netwrk. The key pint is that lw utilizatin may be technlgically inefficient, but it may ften be ecnmically efficient when the ttal system cst is cnsidered. If a newspaper dubles the capacity f the private line between its editrial ffices and the printing plant, the utilizatin rate will drp in half. Hwever, the staff may gain an extra half hur t wrk n the editin befre it ges t press, the half hur that is cut frm the transmissin time f the electrnic layut. Whether that is wrthwhile r nt has t be decided by the managers f the business, and the utilizatin rate is irrelevant. When we see cmpanies rutinely paying fr lightly utilized netwrks, we can cnclude that they d value the ability t send data in high speed bursts, and that shuld guide us in the design and peratin f netwrks. This paper dcuments the lw utilizatin levels f data netwrks mentined abve (and summarized in Table 1). It is likely that sme peple in the cmmunicatins industry understand this already. Fr example, given the aggregate size f private line netwrks (see [CffmanO]), the nly way that the MCI predictin (see Vint Cerf's presentatins at [Cerf]) f data traffic vertaking vice traffic arund the year 2002 can be crrect is if private line netwrks are extremely lightly utilized. Sectins 2 and 3 discuss what netwrks are t be measured, and the units f measurement. Sectin 4 presents data abut switched vice netwrks, t serve as a benchmark in cmparing varius data netwrks. Sectin 5 discusses the backbnes f the public Internet (i.e., thse backbnes that are accessible t general users). Sectin 6 presents data abut sme research netwrks. Sectin 7 is devted t evidence abut utilizatin f private line netwrks. Sectin 8 discusses the reasns that data netwrks are likely t stay underutilized. Finally, Sectin 9 clses with sme cmments and cnclusins. 2. What is t be measured, and why The fcus f this paper is n lng-term average utilizatin f lng-haul lines in the data and vice netwrks, the DS0, T1, T3, OC3, and similar lines that custmers such as ISPs lease frm telecmmunicatins carriers. (Sme carriers, such as AT&T, MCI, and WrldCm, bth wn such lines and use them t ffer Internet services t their wn custmers, and als lease such lines t ther carriers.) 5 Crpratins building private line netwrks and the majrity f ISPs depend n such leased lines, and it is the ecnmics f this business that I wish t explre. I will nt deal with the utilizatin f the fiber netwrk that is used t prvide these prvisined T1, T3, and ther circuits (a fascinating subject in its wn right). I will cnsider nly U.S. data netwrks, althugh there will be sme data abut internatinal links and institutins. The U.S. nt nly accunts fr mre than half f the traffic, but it als has much lwer transmissin csts [ITU, GMLCOBRS]. Therefre its data netwrk behavir is likely t freshadw what will be seen in ther cuntries in the near future, as they expand their telecmmunicatins infrastructure and reduce prices. Only lng distance links will be cnsidered. Fr the vice phne netwrk this will mean nt lking at utilizatin f access links, such as the cpper wire frm a huse t the nearest central ffice r the links frm the central ffice t lng distance switches. Fr data netwrks, LANs (Lcal Area Netwrks) will als nt be cnsidered in detail. They are an imprtant part f the picture, and are discussed at sme length in [Odlyzk1], but in this paper they will be mentined nly briefly. The main reasn fr nt cnsidering lcal links is that their utilizatin patterns differ substantially frm thse f lng-haul facilities. It is widely recgnized that LAN utilizatin is extremely lw. Few peple appreciate just hw lw it is. There are n cmprehensive statistics, but we will cite as ne example the University f Trnt netwrk [Trnt]. The main reasn fr selecting this academic institutin is that its netwrk is unusually well instrumented, with statistics cllected fr all imprtant segments, and displayed with the MRTG prgram f Oetiker and Rand [MRTG]. Trnt is nt prfligate with netwrk resurces, as its Internet link is unusually cngested (as will be discussed later), and s are many f its internal WAN links. Still, the average utilizatin f its 173 Ethernets, during the week ending at 4 pm n Sunday, March 8, 1998, was 1.1%. Only 24 Ethernets had average utilizatin levels ver 2% during that week. Graphs d shw ccasinal spikes in usage (the reasn fr having all that bandwidth), but they tend t be shrt. Even if we take the maximal utilizatin level fr each Ethernet during any 30-minute perid ver that week, and average it ver the 173 Ethernets, we find it is nly 8.7%. The graphs f netwrk usage that are included in this paper are typically fr Sunday and Mnday. The reasn is t shw the different time f day and day f the week patterns f traffic lads n varius netwrks. The implicatins f the similarities and differences in such patterns are be explred at greater length in [Odlyzk1, Odlyzk2]. 6 3. Cnversin factrs It will be cnvenient t state sme cnversin factrs between different units and between the bandwidth f a cnnectin and the traffic carried by that cnnectin. Vice n the phne netwrk is carried in digitized frm at 64,000 bits per secnd. We will be using the cmputer industry ntatin in which Kbps = kilbit per secnd, 1024 bits per secnd. T keep the presentatin simple, we will say that each channel takes 64 Kbps. The inaccuracy this will intrduce is minr. Each vice call ccupies tw channels, ne in each directin, s takes up 128 Kbps f netwrk bandwidth. Thus ne minute f a vice call takes 60*128*1024 bits, r 960 KB (kilbytes). Runding this ff, we get 1 minute f switched vice traffic MB. (Cmpressin can reduce that t a much smaller figure, and is used t sme extent n high-cst internatinal circuits, as well as n sme crprate private line netwrks. As far as the netwrk is cncerned, thugh, it is carrying 1 MB f digital data fr each minute f a vice call.) A T3 (r DS3) line perates at 45 Mbps in each directin, s that if it were fully laded, it wuld carry 90 Mbps. Over a full mnth f 30 days, that cmes t 29 TB (terabytes, bytes). We will say that full capacity f a T3 link TB/mnth. A T1 line (1.5 Mbps) is 1/28-th f a T3, and we will say that full capacity f a T1 link TB/mnth. 4. Switched vice netwrks It is interesting t nt nly estimate utilizatin levels f varius data netwrks, but als t cmpare them with the circuit switched netwrk. The bk [Keshav] is an excellent surce that cntrasts the technlgies invlved in these types f netwrks. Hwever, n cmprehensive descriptin f hw they are used appears t exist. Figure 1 shws the typical traffic pattern n U.S. switched vice netwrks. It is derived frm Fig f [Ash]. This graph aggregates all the phne calls ver the fur time znes f the cntinental U.S., as well as the cmparatively small number f calls t Hawaii, Alaska, and ther places. (Fr 7 mre data, including calling patterns in smaller regins, see [Ash].) Vice netwrks, such as that f AT&T, are engineered t prvide a lw-cst slutin t all nrmal demands. This means that many calls may get blcked in cases f an earthquake, say, but even peak hur demands during the busiest days, such as Mther's Day r the Mnday after Thanksgiving, are accmmdated. Fr example, t cite a small sample f the data in [Ash], n Mnday, Dec. 2, 1991, which was the busiest day fr the AT&T netwrk until then, f millin calls, nly 228 were blcked n intercity cnnectins. In spite f this, the average utilizatin f lng distance links in the switched vice netwrk is clse t 33%, as is explained in [CffmanO], based n data frm [Ash]. This efficiency cmes frm careful engineering (using techniques such as RTNR, Real Time Netwrk Ruting [Ash], that rute calls between New Yrk City and Philadelphia thrugh Chicag when spare capacity is available n thse rutes), frm the smther and mre predictable nature f vice traffic in general, and the predictable grwth in demand fr vice services. An imprtant cntributr t the high average utilizatin f vice netwrks is the sharing f this netwrk amng several classes f users with different calling patters, a pint explred at greater length in [Odlyzk1, Odlyzk2]. Average utilizatins are far lwer if ne cnsiders the entire telecmmunicatinis netwrk. There are extensive circuits that exist t prvide service in case f fiber cuts and similar utages. These circuits have large capacity, but they are used t prtect data circuits as well as vice lines, and are utside the scpe f this paper. 5. The public Internet The Internet is slw, as anyne wh surfs the Web can attest. Hwever, it has prved impssible s far t prduce a simple descriptin f where the prblems lie. (Fr the mst thrugh statistical study f Internet perfrmance currently available, see [Paxsn].) Many f the prblems are with the servers. Hwever, the general impressin is that the backbnes are seriusly cngested. This view is supprted by studies f cmparative backbne perfrmance, which d shw substantial differences in perfrmance amng differe
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks