David Roberts Publication Rev. Amendment Issue Date

The Datasheet Archive - 100 Million Datasheets from 7500 Manufacturers.

David Roberts Publication Rev. Amendment Issue Date

Datasheet Thumbnail

Top Searches for this datasheet

Effect Local Buffer Memory Size FDDI Throughput
David Roberts
Publication
Rev. Amendment
Issue Date
1992 Advanced Micro Devices, Inc.
16294
1/92
Effect Local Buffer Memory Size FDDI Throughput
David Roberts
ABSTRACT
effect size local buffer packet memory throughput FDDI adapters examined. analysis focuses transmit buffer size covers both loaded unloaded FDDI ring cases. hypothetical EISA FDDI adapter design used vehicle demonstrate throughput characteristics different local buffer memory sizes. Calculations done size buffer with respect latency EISA card target system configuration with respect amount effective bandwidth that card expect receive. demonstrated that sizing buffer based these methods results poor throughput when FDDI ring loaded. solution performance problem, suggested that buffer sizing based TTRT value ring will result adapter that capable realizing high throughput. Finally, SUPERNET2 FDDI chipset examined show that able meet diverse buffer sizing needs.
high effective throughput nodes that network. this FDDI access protocol designed insure that network efficiency, defined ratio obtainable bandwidth under high load Mbps maximum bandwidth, would high. Additionally, designers tried ensure that obtainable bandwidth fairly allocated among nodes that were competing Fairness achieved through FDDI access protocol. Although FDDI access protocol provides fair distribution large, obtainable bandwidth, nodes able realize full throughput available them, depending their design. Poorly designed nodes even limit total network performance destroying available bandwidth needlessly. This paper will demonstrate local buffer memory size chosen adapter cards affect throughput achievable that card network. particular, will shown that transmit buffer size that takes into consideration value TTRT will lead optimum performance adapter over both loaded unloaded network system conditions. following section defines some terminology acronyms that used subsequent sections. Section short overview FDDI, basic access protocol, some parameters which govern performance. Section describes some system parameters that must understood when designing FDDI node. Section uses FDDI EISA card design example shows parameters affect performance card under various system network loads. section demonstrates that correctly tuned buffer memory will make card more robust when additional stress applied final section, analysis from previous sections compared with buffer memory size offered AMD's SUPERNET FDDI chipset. comparison shows that SUPERNET able meet variety buffer sizing needs methodologies.
INTRODUCTION
Fiber Distributed Data Interface (FDDI) megabits second (Mbps) local-area network (LAN) standard developed standardized American National Standards Institute. FDDI standard designed meet greater bandwidth requirements next generation computing information equipment. FDDI developed allow rapid transmission computer data well synchronous information, such video, using ring topology similar 802.5 (Token Ring) standard. FDDI allows stations separated maximum cabling distance between stations, with total cabling length limited FDDI represents full order magnitude increase over bandwidths offered based networking standards such 802.3 (Ethernet) 802.5. distances between stations high transmission bandwidth made possible because standard based fiber optic cable physical media (hence FDDI). Schemes transmission same data rates, with station separation distances limited meters, using shielded unshielded, twisted pair, copper wiring being investigated ANSI X3T9.5 committee offer lower cost medium when great distances needed. original FDDI design objectives that high bandwidth FDDI should translated into
Publication Rev. Amendment Issue Date
DEFINITION TERMS
This section defines some acronyms, terms, variables that will used later sections. Although many these defined various other locations this paper, this section provides convenient reference point those that need
16294
1/92
1992 Advanced Micro Devices, Inc.
Effect Local Buffer Memory Size FDDI Throughput
access delay that node experience when wishes transmit network. BWnet: bandwidth FDDI network. This constant equal Mbps. ring latency FDDI ring. DAS: Dual Attachment Station. FDDI node which connects both counter rotating rings FDDI network. efficiency that system able achieve. EISA: Extended Industry Standard Architecture. extension original bus. EISA features wider (32-bit) address data busses well faster transfer protocols. EBWsys: effective bandwidth system bus. FDDI: Fiber Distributed Data Interface. Mbps ANSI standard based fiber optic media, ring topology, token passing access protocol. ISA: Industry Standard Architecture. expansion architecture used PC-AT class personal computers. arbitration latency bus. LAN: Local Area Network Nsys: number bytes that system transfer during given time interval. SAS: Single Attachment Station. FDDI node which attaches only ring FDDI network. T4500: amount time necessary transmit 4500 bytes data FDDI network. interval time refill local buffer with data across system bus. THT: Token Hold Time. timer this described positive-valued down counter this paper. value computed TTRT TRT. positive number, then token "early" relative negotiated TTRT value. this case, asynchronous data transmitted node this token rotation. negative, zero, then there asynchronous bandwidth available node token must passed after transmitting synchronous data that node have. throughput that system able achieve. TRT: Token Rotation Time. actual time seen between successive tokens given node. timer this described positive-valued counter this paper, which differs from other descriptions many actual implementations. feel that this makes timer system little easier understand.
TTRT: Target Token Rotation Time. time negotiated nodes FDDI ring when initialized which used bound time that node between successive tokens. value positive. size local transmit buffer memory. 802.3: Mbps IEEE standard based topology utilizing CSMA/CD medium access protocol. 802.5: Mbps IEEE standard based ring topology token passing access protocol.
OVERVIEW FDDI
following simple description FDDI some features access protocol. following meant complete description FDDI. Rather, just introduces some actions taken FDDI designers insure high efficiency FDDI protocol, more important this discussion, introduces concepts Target Token Rotation Time, Token Rotation Time, Token Holding Time. These will important following sections. FDDI basic topology counter rotating rings. rings used carry data other provides redundancy event failure link between stations. event failure, rings will wrap either side failure into large ring, restoring service. Stations which connect both rings called single attachment stations (SAS). They connect only primary ring through concentrator, which may, not, connect secondary ring. Although both ring topology FDDI token access ring similar 802.5, there many differences actual access protocols between 802.5 FDDI. addition providing high, bandwidth, designers FDDI felt that important insure that bandwidth available network usable individual nodes, even when network busy. Bandwidth consuming collisions avoided using token access protocol. special packet data, called token, passed from node node around ring. Nodes needing transmit data hold token when they receive called "capturing token", transmit some data, then pass token next station downstream. This fundamental concept same both 802.5 FDDI. 802.5 protocol, frames that transmitted network must removed from network transmitting station. Mbps version standard, station finishes transmitting, must wait, transmitting nothing, until data that onto ring returned full circle. Following this, then transmit token next station. This scheme wastes
Effect Local Buffer Memory Size FDDI Throughput bandwidth because there dead time ring while station waits data return fact, this problem later corrected when Mbps version standard developed. FDDI, designers allowed early release token. Nodes must still strip their transmitted frames from network when they receive them again, they transmit token next node soon they done transmitting their data. next node turn then capture token start transmitting immediately. Because this, there little dead time loaded ring. FDDI designed carry synchronous data, such video, addition computer data. available bandwidth network split into types, synchronous asynchronous. Nodes requiring guaranteed bandwidth synchronous transmission mode; every token rotation node captures token transmits guaranteed length time. This time negotiated among nodes wishing synchronous transmission mode, depending their synchronous bandwidth requirements. Nodes that require guaranteed bandwidth asynchronous transmission mode FDDI access protocol. When FDDI ring initialized, Target Token Rotation Time (TTRT) negotiated among nodes. TTRT suggested time that token should take circulate around ring when ring loaded. most, token will take 2TTRT circle ring (this will occur when ring previously quiet nodes begin transmitting both synchronous asynchronous data their bandwidth limits). TTRT represents bounded time quantum between transmissions that nodes require their bandwidth needs. token circles ring, each node watches successive token arrival times token early late arriving, compared TTRT. node measures interval between token arrivals with timer called Token Rotation Timer (TRT). Another timer, Token Holding Timer (THT), defined TTRT TRT. positive number, then represents amount time that token early, relative negotiated TTRT. When token early, node opportunity transmit asynchronous data. negative, then token late asynchronous data transmitted this token rotation. token must passed immediately node's downstream neighbor. node always transmits synchronous data whether token early late. After transmitting synchronous data that ready, node starts decrementing THT. While greater than zero, node transmit asynchronous frames. When expires reaching zero, node must pass token downstream neighbor. node runs data transmit before expired, should pass token. node holds
when data transmit, bandwidth will wasted. TTRT, TRT, important understand because they affect Network Latency parameter described next section. more information about FDDI operating protocols, [4].
IMPORTANT PARAMETERS
This section introduces some parameters that important FDDI node's performance. These parameters assume adapter card model, shown Figure should noted that many "mother board" designs, although they don't separate adapter card, also fall into this model. assumed this model that host processor ultimate recipient data that transmitted received FDDI adapter. Therefore data must moved from system memory whenever packet transmitted received.
System FDDI Ring
System Memory
Adapter Card
FDDI Adapter Card
Host Processor
Adapter Card
16294A-001A
Figure Adapter Card Model
Common Parameters
parameters that most often used help predict local buffer requirements system performance FDDI node peak bandwidth, latency. These described below.
Peak Bandwidth peak bandwidth peak theoretical rate transfer between system memory FDDI adapter. peak bandwidth less than FDDI maximum transfer rate, then system will bottleneck data transfer node node will
Effect Local Buffer Memory Size FDDI Throughput tocol allows transmit. case FDDI, this corresponds interval when node able find capturable token. interaction between buffer memory size this parameter will shown later sections. node transmitting asynchronous data must pass token because expires, access delay amount time must wait before capture another usable token begin transmitting again.
never able sustain full FDDI bandwidth, even both FDDI ring system unloaded. Further, whenever node receives data from network, will unable move packets into system memory faster than incoming network reception rate. long stream packets directed node, node overrun, leading data loss.
Latency This parameter describes length time that FDDI adapter must wait access system after decided that will arbitrate This important parameter because most networks inherently asynchronous with respect host system. When inbound packet arrives adapter transferred system memory when transmit opportunity occurs cannot predicted. When these events happen, adapter must access system memory either transfer inbound packet transfer outbound data out. adapter must have enough buffering capability locally overcome latency during time that waiting.
CASE STUDY EISA
following section uses Extended Industry Standard Architecture (EISA) short example various parameters that were presented previous section affect amount local buffer memory that should included FDDI adapter design. EISA developed consortium vendors 32-bit extension Industry Standard Architecture (ISA) bus, which used PC-AT class personal computers. EISA supports 32-bit address data paths, software configuration cards power high speed transfer modes with peak bandwidths MB/s. Additionally, upward compatible with older cards. more information about EISA, [2].
Frequently Overlooked Parameters
following parameters also important deciding much buffer memory needed local FDDI adapter. Although these parameters often overlooked, they very important FDDI designs. high data rate FDDI along with timed token access protocol demand that these parameters understood adapter card designers.
Sample Configuration
During this example, imagine that responsible design EISA FDDI adapter card. adapter card been targeted server product model configuration contains following cards: 802.3 Ethernet Master SCSI Master FDDI Master, which designing wondering cards masters none slaves. will assume that cards system high performance EISA cards that need master themselves either manage advanced data structures like descriptor rings, case networking cards, download command scripts, case SCSI card. This complex manipulation cannot done card itself unless master. EISA preemptable EISA specification defines exactly long card remain once been preempted. case master, specification allows card continue stay BCLKs after preempted. will assume that since networking cards asynchronous nature, they will always stay BCLKs after being preempted reduce chance under overrun. will also assume that since SCSI handshaken protocol cannot overrun FIFOs local buffers, that SCSI card will only remain BCLKs after preempted. This allows
Effective Bandwidth Effective bandwidth refers actual transfer rate that network adapter into host memory over extended period time. unloaded then effective bandwidth same peak bandwidth. other devices using over period time that FDDI adapter transferring data, then effective bandwidth allocated FDDI adapter will less than peak bandwidth.
Some busses allow designated devices have priority higher than that other devices. arbitration scheme such that once FDDI adapter arbitrated bus, that stay resident long would like (i.e. adapter preemptable) then effective bandwidth available that card equal peak bandwidth. calculation effective bandwidth takes into account above parameters, transfer rate latency, third parameter, transfers arbitration cycle.
Network Access Delay This parameter similar latency that defines time difference from when FDDI adapter knows that packet transmit when access pro4
Effect Local Buffer Memory Size FDDI Throughput more bandwidth those cards that need when system becomes busy. Finally, will assume that FDDI card bursting master. other words, FDDI card able achieve full EISA peak bandwidth bytes/s during time that transferring bus. This peak bandwidth this card, defined Section 4.1.
shown table completeness, since there slaves system configuration, arbitration slots contribute nothing total arbitration latency FDDI card. more information about EISA arbitration scheme, [2]. Table EISA Arbitration Latency Calculation FDDI Card Device FDDI Master Refresh Refresh 802.3 Master Refresh Refresh SCSI Master Refresh Refresh FDDI Master Hold Time (µs) 10.6 Total Time (µs) 10.3 10.3 10.3 20.9 20.9 22.2 31.2 31.2 32.5 38.3 38.3 38.3 47.3 47.3 48.6 48.6 Notes Release
Calculation Latency following calculation FDDI card's latency parameter. This maximum time that card expect wait bus, this system. following calculations, same assumptions used EISA specification calculate system latencies will used here, those that applicable. assumptions taken from Section 2.9.2 EISA Specification. These assumptions are:
EISA bus. releases within BCLK periods plus completion time locked cycle) after preemption occurs. Masters release within 10.6 BCLK periods plus completion time final cycle) after preemption occurs. controller (programmed BLOCK DEMAND mode) releases within BCLK periods plus completion time final cycle). refresh cycle takes CPU, channels, masters reassert their request signal immediately after relinquishing after preempt. case, there cards that slave system; cards masters. SCSI card, however, will number that specification uses preemption time because SCSI card stays BCLKs, like channel would. card will arbitrate master, though, should. fact that devices will reassert their requests immediately after giving indicative busy system. Thus latency calculating worst case this particular configuration cards. following table, Table shows sequence events that occur after FDDI card relinquishes control rearbitrates order arbitration specified EISA specification. Simply, arbitration round-robin between three major groups: masters, channels, refresh. master category, arbitrates every other cycle. masters rotate through remaining slots this group. Refresh arbitrates once every arbitration slots slaves
Total
Total
Total
Total
Total Grant
column labeled "Total Time" shows accumulated time since FDDI card released until gets back. Since card rearbitrates immediately after relinquishing control, from table that 48.6 maximum latency FDDI card this system configuration. latency could much longer system that more cards very busy, much less this same system busy all.
Calculation Effective Bandwidth Although peak bandwidth EISA bytes/s, this bandwidth only achievable while card bus. bandwidth available FDDI card over period time, effective bandwidth, much less because demands placed CPU, other cards, refresh mechanism. Knowing effective bandwidth will help size local buffer memory FDDI adapter. effective bandwidth FDDI card this configuration calculated from peak bandwidth, residency time, latency.
Effect Local Buffer Memory Size FDDI Throughput rest packet from system memory card's local buffer. Given this scenario, much buffer memory must card that transmission doesn't underrun when engine tries restart arbitrating (i.e. overcome latency)? question restated following way: many bytes network transmit before adapter gain control start refilling buffer again? know that network data rate FDDI Mbps, which equivalent 12.5 bytes/s, 11.92 MB/s 1,048,576 bytes). During adapter's arbitration latency 59.2 calculated previous section, adapter will able transmit (12.5 bytes/s) (59.2 bytes. Thus, need have least bytes data card, ready FDDI adapter transmit, that card rearbitrate have packet buffer underflow during that time. shall see, amount memory necessary overcome latency minimum amount buffer memory necessary design.
peak bandwidth bytes/s. With clock rate MHz, this works 32-bit word being transferred each clock cycle, which indeed, EISA burst protocol designed function. residency time BCLKs after card preempted, which will assume happens same time that card takes control bus. Thus each time that card starts transferring data, gets preempted must within BCLKs. word data transferred each clock while card bus, this amounts words, bytes, data being transferred each time card gains control. Assume that residency period 10.6 specification says masters that staying BCLKs after being preempted. Adding residency time arbitration latency gives whole arbitration cycle period. This 10.6 48.6 59.2 Thus, every 59.2 bytes data transferred card. effective bandwidth card then: (256 bytes/arb cycle)/ (59.2 µs/arb cycle) 4.32 bytes/s, 4.12 MB/s 1,048,576 bytes). Since calculation effective bandwidth dependent arbitration latency, important note that 4.12 MB/s effective bandwidth FDDI adapter this configuration under this load, only. another card added system, increasing load, then effective bandwidth adapter will decrease further. load system decreases because more SCSI 802.3 activity present, then effective bandwidth available FDDI adapter will increase.
Overcoming Limited Effective Bandwidth effective bandwidth available card fill buffer larger than emptying rate FDDI network, then system latency dominant issue determining size local buffer memory. system will able supply data faster than network will able transmit thus there will rarely underrun.
practice, however, often very easy show that effective bandwidth available card fall below network transmission rate 12.5 bytes/s. This especially true when card interfaces slow bus, when fast heavily loaded. case EISA server example, peak bandwidth bytes/s, which more than double network transmission rate. spite this, demonstrated that when configuration loaded, effective bandwidth available FDDI card drops 4.32 bytes/s. When effective bandwidth available fill buffer memory less than network transmission rate, card unable totally refill buffer memory when gains access bus. Thus, memory sized simply overcome latency, buffer might underflow during next arbitration period. this situation, question becomes: much buffer memory needed prevent underrun because limited system fill rate? answer depends packet that transmitting packet size less than bytes, then already into byte memory will never underrun when transmitting.
Minimum Buffer Size Prevent Transmit Buffer Underrun
following sections describe calculations necessary size buffer memory EISA card such that, during transmission packets, buffer underrun cannot occur. cases examined: first overcome maximum EISA latency card expect, second overcome limited effective bandwidth available card transfer data. This discussion applies both synchronous FDDI transmission mode well asynchronous mode.
Overcoming Latency imagine following situation: formatted outgoing FDDI packet which wishes transmit onto FDDI network. preparation transmitting packet, told card packet into card's local buffer memory. engine starts continues until card's local buffer full, then stops. When useable token captured card, starts transmit packet onto network, emptying buffer memory. buffer memory empties, engine restarts, transferring
Effect Local Buffer Memory Size FDDI Throughput Additionally, since peak bandwidth EISA available once gain access bus, card able download rest packet local buffer arbitration cycle (i.e. rest packet bytes), packet will never underrun. worst case, after card been granted bus, byte buffer memory will empty. card will have just transmitted last byte from buffer onto network. card fill memory bytes/s bytes. Since bytes/s fill rate larger than 12.5 bytes/s exit rate network, amount data buffer will increase during this time, packet won't underrun. Remember, however, that adapter designers, would like ensure that maximum size packet will able transmitted under variety circumstances without underruns. case FDDI, need able transmit 4500 bytes without underrun. wanted take easy out, would simply 4500 byte memory local adapter transfer entire packet buffer before started transmitting data network. practice this inefficient because amount storage that actually needed probably less than 4500 bytes. Imagine that about transmit 4500 byte packet network. bytes data reside adapter local buffer, with 4500 bytes residing system memory, waiting transferred adapter. This starting configuration system. Just adapter starts transmitting onto network, engine adapter will start transferring remaining bytes from system memory local buffer memory. size buffer optimally, based effective bandwidth, following condition should hold transmitting packet: just network about transmit final byte packet onto network, engine must have just transferred last byte across system into local buffer. this condition met, live ideal world, then adapter will able transmit last byte without problem packet will underrun. task compute given these conditions. Solving takes four steps. Compute time that takes adapter transmit 4500 bytes network. Call this time T4500. fact that byte bandwidth FDDI network (BWnet) 12.5 bytes/s Mbps/ bits/byte). Compute number bytes that transferred over system during time T4500, given effective bandwidth (EBWsys) 4.32 bytes/s. Call this amount data Nsys. Assert that Nsys 4500 Solve Step
T4500
(4500 bytes) BWnet (4500 bytes) /(12.5 bytes/s) 10-6
Step
Nsys
EBWsys T4500 (4.32 bytes/s) (360 1555 bytes
Steps
Nsys
4500
4500 Nsys
4500 1555
2945 bytes 2.88 1024 bytes) Step always gives constant value because maximum FDDI frame size FDDI transmission rate fixed FDDI specification. only independent variable calculations effective bandwidth system bus. Thus, these four steps reduced following simple formula:
(4500 bytes) [EBWsys (360 10-6
From above calculations that must have least 2.88 local buffer memory FDDI adapter able support transmission 4500 byte packet without underrun, given effective system bandwidth 4.32 bytes/s. see, this much larger number than bytes that needed just prevent underrun because system latency. this case, effective bandwidth dominant factor computing size buffer. Further, shall that 2.88 simply amount memory that needed support correct operation under above system load. following sections will show that overall network throughput will extremely larger memory needed support high throughput. Furthermore, important note that above analysis transmit path only. 2.88 value computed above does take into account memory that needed buffer incoming frames. shown, with analysis similar above, that minimum amount memory that needed buffer incoming frame, assuming same 4.32 bytes/s effective system bandwidth, also 2.88
Effect Local Buffer Memory Size FDDI Throughput
receive case, system will have just transferred 1555th byte incoming 4500 byte packet from local buffer system memory when last byte packet copied from network buffer FDDI adapter. first packet immediately followed second, however, second packet will overrun buffer will have dropped because engine will able empty buffer fast enough make room second incoming frame. handle case many frames arriving quick succession, buffer size will have increased beyond 2.88 correctly size this buffer expected reception traffic load needed well. This left another analysis. rest this paper focuses transmission only. Some readers take exception fact that this example assumes that EISA always devices that connect CPU, refresh mechanism, three cards. general system would probably quiet much time. FDDI card were access randomly, would find very latency average would able sustain close MB/s peak bandwidth that EISA allows. server configuration, however, often case that disk usage closely coupled with network activity. Generally client workstation generates series disk read write requests over network which leads server activity process packets, which turn leads disk activity service requests. This turn leads more activity more network activity requests fulfilled. Although average system quiet, will become busy bursty manner which could lead contention effects which have predicted previous sections duration burst. client workstation that performing file transfer experience similar activity. Although analysis assumes that single 4500 byte packet will transmitted, argument also holds multiple smaller sized packets whose total number bytes 4500. analysis just computes minimum buffer size support 4500 byte transmission. fact, many network protocols, such Internet Protocol (IP), default packet sizes smaller than 4500 bytes. does place limit this, however, many implementations starting packet size when FDDI recognized underlying network. This reduces packet software processing overhead. analysis does include delay time system transmission that associated between additional packets, however. multiple smaller packets used, buffer size needed support 4500 byte transmission should larger than amount that calculated here support additional overhead time.
Throughput Calculations Minimum Buffer Size
Given buffer size calculated section 5.2, 2.88 this section examines performance characteristics card FDDI side interface. Specifically, this section investigates FDDI throughput that card will able achieve when network both unloaded loaded. Although above discussion applied synchronous transmission well asynchronous transmission, this section will only address asynchronous transmission. throughput synchronous transmission guaranteed negotiation synchronous bandwidth allocations among nodes.
Ring Model Assumptions following discussion assumes configuration FDDI ring which EISA card attached. configuration single attach stations (SAS) fiber ring will used. This configuration chosen because identical "typical" configuration, used Jain [1]. Jain states that these numbers based intuitive feeling what typical ring would look like based survey actual installations. This configuration could represent floor office building. Twenty offices, using SAS, located floor would require worth cable fiber ring.
will assume that TTRT been negotiated among stations. Jain found that this value provides high ring efficiency maximum access delay over wide range ring sizes loadings [1]. Additionally, will assume that speed light through fiber 5.085 µs/km that delay that suffers when passing through station order These same assumptions used Jain. Given ring configuration above assumptions, able compute ring latency. ring latency describes amount time that takes data circle ring return transmitting station. Calculating ring latency very straight forward. simply travel time light representing each through total fiber total delays through each node ring. case, ring latency,
(5.085 µs/km) stations) µs/station)
section computed minimum amount local buffer that required insure that 4500 byte FDDI packet will suffer transmission underflow, given effective system bandwidth 4.32 bytes/s.
Effect Local Buffer Memory Size FDDI Throughput During analysis, assumed that avoid underflow, last byte packet would have transferred over system just before required transmitted network. Thus, after transmitting 4500 bytes network, FDDI adapter will forced pass token downstream neighbor because does have more data ready transmit. adapter began transmit another packet, would immediately underrun. held token, waiting more data transferred over system bus, would waste FDDI bandwidth. following throughput analysis, therefore, will assume that after each 4500 byte block transmitted, node must release token back FDDI ring.
card restarted transfer rest 4500 bytes first packet local buffer. T4500, entire 4500 byte packet been transmitted network. local buffer memory empty, thus FDDI card reissues token ring begins refill local buffer with data from second packet. effective system bandwidth 4.32 bytes/s, 2.88 local buffer will fill This time interval much larger than that takes token circle ring. Thus token will circle number times while buffer being filled. time T4500 buffer full node waits token pass that node capture This assumes that 4500 byte packet transmitted. smaller packets used, token able captured earlier. this point, token passes captured. time difference from time when node starts waiting token time which captures called network access delay (See Section 4.2). When ring unloaded, maximum access delay seconds, since token travels around ring maximum rate node transmitted several token rotations. Again, 4500 byte frame transmitted token released because there more data local buffer after 4500 bytes have been transmitted. throughput achieved card above example easily calculated. single frame transmitted every T4500 seconds. even multiple ring latency, then missing fraction also added make total even multiple case, which little more than 17D. Thus, T4500 fixed network rate period each transmission cycle then 1.08 throughput (4500 bytes)/(1.08 bytes/s. Although FDDI bandwidth available card transmission 12.5 bytes/s, throughput card limited effective system bandwidth 4.32 bytes/s. card will only able achieve maximum throughput equal this rate. Given that being limited 4.32 bytes/s, bytes/s that actually able achieve very efficient. more exact, (4.2)/(4.32) efficient. case unloaded ring, example card performs very well.
Unloaded Ring Throughput Consider unloaded FDDI ring, configuration given above, with example EISA card attached nodes. time EISA node becomes active wants transmit number 4500 byte packets onto network. following figure, Figure shows sequence events that occur. Time flows from figure toward bottom. reference, other node ring (any node, doesn't matter which) shown. token arriving node ring being issued following transmission shown bold horizontal line. bold vertical line indicates that token being held that frame being transmitted particular node.
EISA Node Other Node Token 4500 byte frame
T4500
Token rotates several times while buffer fills. Time
16294A-002A
Figure Unloaded Ring Event Diagram Previous time token circles ring, endlessly, with rate because ring unloaded. During this time system fills 2.88 buffer with first part first packet that must transmitted. time node captures token begins transmitting first packet. same time, engine
Loaded Ring Throughput let's card performs when ring slightly loaded. Assume that node begins transmitting, other node ring begins transmitting. Also assume that other node capable utilizing
Effect Local Buffer Memory Size FDDI Throughput EISA Node
FDDI bandwidth, itself. That once other node captures token, will continue transmit until Token Holding Timer expired. will release token before FDDI access protocol, because fair, will make sure that card opportunity utilize half 12.5 bytes/s FDDI bandwidth. following figure, Figure illustrates sequence events that occurs when nodes both start transmitting. time both node other node start trying transmit. EISA-based node captures token first starts transmitting first 4500 byte frame. When transmitted 4500 bytes, cannot transmit more frames because local buffer empty, releases token starts filling buffer. other node wishing transmit captures token. node sets amount time that token early. amount, this case, TTRT T4500 seconds. node starts decrementing begins transmission. After TTRT T4500 seconds, other node's expires, node stops transmitting. then releases token ring. token arrives back EISA node. Unfortunately, token right time. Token Rotation Timer, which node been incrementing since token last arrived, exactly equal TTRT that negotiated when ring initialized. Because token early, node cannot transmit asynchronous data this time. EISA node passes token. other node again captures token, because early, sets amount time that token early. this case, amount time T4500 seconds. Examine diagram this node begins transmitting data decrementing THT. other node's expires. transmitted 4500 bytes worth data. releases token back ring. node captures token transmits more 4500 byte packet. then releases token because more data ready transmit. begins refill buffer. other node captures token. sets TTRT 2T4500. After transmitting this amount time, releases token once again. EISA node captures token begins transmitting. After transmitting another 4500 bytes, releases token back network. starts refill buffer. other node captures token transmits 4500 bytes, after which expires releases token. node captures token. transmits 4500 bytes then releases token. Some readers notice this slight mistake analysis.
Other Node
Token
T4500
4500 byte frame
TTRT TTRT T4500
Time
T4500
T4500
TTRT 2T4500
T4500
T4500
T4500
TTRT 2T4500
16294A-003A
Figure Loaded Ring Event Diagram
Effect Local Buffer Memory Size FDDI Throughput fact, EISA node should able capture token this time. token only been gone T4500 seconds. know, from analysis about minimum buffer sizes preventing underrun, above, that during T4500 token absence, node only transfer about 1700 bytes across system bus. know that must have 2945 bytes buffer prevent underrun packet. Because this, node should pass token because ready transmit. simplify analysis, however, will allow node capture token this point. data composed several smaller packets, then station might able capture token, although would able transmit full 4500 bytes. analysis throughput value that calculated will higher than actual value, because both letting node capture token before should able assuming that able transmit full 4500 bytes. After steps through sequence steps through becomes repeating cycle. period this cycle TTRT T4500. During this time sequence, EISA node transmits twice, steps total 9000 bytes. throughput node
will experience transmit underrun while system busy. this end, designed card with minimum amount outbound buffer memory support transmission 4500 byte frame without underrun. When FDDI ring unloaded, that card able perform limit effective system bandwidth. design very efficient converting both available FDDI bandwidth system bandwidth into data throughput. When ring became loaded with just other node, however, throughput that card able achieve dropped dramatically. Although both effective system bandwidth available FDDI bandwidth were bottlenecks, card only able achieve throughput about effective system bandwidth. Let's short analysis figure performance low. Consider loaded ring example, above. After capturing token transmitting 4500 bytes, card releases token back network. card's 2945 byte buffer empty begins refill with next packet. When buffer full, card waits capturable token pass network unloaded, then time that card will wait, network access time, less than equal however, network busy, then card have wait long time. Jain shown that worst case, node wait long 1)TTRT seconds, where number nodes that actively transmitting FDDI ring time. This occurs when nodes being "greedy" transmitting until their THTs expire. important note that node guaranteed token 2TTRT seconds. This must case those nodes that have synchronous data transmitted. That token capturable asynchronous transmission, however, because value time greater than equal TTRT. token step loaded ring diagram, above, example such token. Jain's equation gives upper bound node find token that capturable. Some readers have also noticed that case only active node, equation predicts maximum value access delay, compared value just that stated couple paragraphs ago. these different? reason that Jain's formula predicts maximum delay with assumption that node transmits full time then immediately tries recapture token. this case, first token that passes node will uncapturable because node transmitted full THT. Draw small diagram like previous ones, this still unclear. case, however, node only transmitting 4500 bytes. only didn't transmit full time, also token spin around network sev-
(9000 bytes)/(TTRT T4500) (9000 bytes)/(8 (9000 bytes)/(8.4
1.07 bytes/s Given 4.32 bytes/s maximum band-width, efficiency FDDI card with other node transmitting FDDI ring
(1.07 bytes/s)/(4.32 bytes/s)
card only able transmit one-fourth data that should able Remember, also, that this efficiency higher than what card should actually able accomplish because simplification made step analysis. important note that neither network system limiting card. network access protocol guarantees that card opportunity transmit throughput almost 6.25 bytes/s. have previously shown that effective system bandwidth 4.32 bytes/s. card only able attain throughput 1.07 bytes/s?
Problem Analysis
Problem analysis given section based assumption that card designers concerned with correct operation card during loaded situation. That simply want insure that card
Effect Local Buffer Memory Size FDDI Throughput node after suffering rotation delay around ring. Since TTRT, bounded approximately TTRT. Thus, local buffer should larger than number bytes that transmitted network during TTRT seconds. buffer larger than this, then extra data won't transmitted during token capture, storage wasted. Given that buffer should larger than size predicted underflow calculations, above, smaller than number bytes that transmitted network during interval TTRT seconds, should buffer
eral times while refilled local buffer memory. Because these things, next token pass node guaranteed capturable, leading maximum delay just loaded ring example, above, that maximum access time node TTRT 2T4500. Using value TTRT that assumed above, compute exact value network access delay situation that described:
TTRT 2T4500 2(360 7.28
Remember that this value only takes into account other node FDDI ring being active. many nodes active, network access delay much longer. stated earlier, time that takes fill 2945 byte local buffer over system only After filling buffer node could possibly wait 7.28 before gets token again. During this time period, system transferring more FDDI data that stored system memory! There "dead time" system bus, with respect FDDI adapter (i.e. other cards could using bus, isn't totally dead). This dead time caused simply fact that buffer card full.
Calculation Buffer Size Assure Maximum Usage Available Token Holding Time
tune size buffer, based TTRT value, will argument that similar used calculate buffer size based effective system bandwidth. That will assume that only part total amount data that wish transfer will actually present card when card starts transmitting data network. remaining portion data will supplied system, across EISA bus, during transmission data network. formula just:
[TTRT/(80 ns/byte)] [EBWsys TTRT]
first term formula calculates total number bytes that wish transmit during TTRT period. Eighty nanoseconds byte rate data transmission FDDI network. second term computes much data system supply card during time that node transmitting full TTRT seconds. difference these values number bytes that must card prevent underflow this block data during transmission. following calculation shows much memory needed example configuration, described above.
Solution achieve throughput that approaches effective system bandwidth, system needs transfer data during whole 7.28 network access delay. When this happens, will transferring data constantly, both when node captured token when waiting, throughput card will equal effective system bandwidth. achieve this, buffer card must made larger.
After reading above paragraph examining Jain's formula maximum access delay, might believe that card needs have buffer that able sustain very large network latency. Further, size this latency defined size ring, number nodes transmitting ring, negotiated TTRT value, which vary greatly. Computing size this buffer would seem difficult problem card designer. Fortunately, there another constraint system THT. When node captures token, only transmit limited time. duration current value, which equal amount time that captured token early arriving. most, equal TTRT This occurs when token transmitted other node wishes able capture token. token simply returns transmitting
ns/byte)] [(4.32 bytes/s) ms)] 100,000 bytes 34,560 bytes
65.44 bytes This value much larger than 4500 bytes that might assume that should size buffer were simply trying prevent underrun. important note that above calculation based TTRT value This value chosen example because Jain shown that optimal over great variety ring sizes configurations. Some FDDI rings higher value, however. Usually
Effect Local Buffer Memory Size FDDI Throughput value order tens milliseconds. TTRT used, instance, then equation predicts buffer size course, this discussion neglects effect that software play sizing buffer. Realistically, protocol stack running host processor will impose some limitations amount data that will ready transmitted time. Many TCP/IP implementations, instance, sliding window protocol which limits amount unacknowledged data eight packets. Although default packet data size bytes, implementations using FDDI starting size their packets based FDDI maximum packet size 4500 bytes. This reduces number packets transmission, thus total software overhead block data. common packet data size these implementations Using this packet data size, along with eight outstanding packets, show that local transmit buffer size about probably that needed. software will have more than this amount data transmitted time, sizing buffer beyond this point will bring minimal performance gains. important note that other implementations other protocols have different requirements. final size should take into account present future trends protocols size buffer.
transferring data, throughput FDDI node limited effective bandwidth FDDI adapter system bus. adapter able achieve almost 100% data rate that system able supply, 12.5 bytes/s FDDI data rate limit.
Loaded Ring Throughput case lightly loaded ring, where network access delay less than amount time required system fill buffer, same sequence events occur unloaded ring example. system never stops transferring data card. Therefore, adapter able achieve throughput approaching 100% system effective bandwidth.
Consider, however, case very loaded FDDI ring, where network access delay much greater that interval time required system fill buffer. this case, system will fill buffer while node waits capture token will stop transferring more data. system will idle with respect FDDI adapter while node waits. When token finally captured node, however, node will transmit data TTRT seconds, until preempted expiration. this case, FDDI bandwidth allocated node FDDI access protocol, based network load, decreased point where less than system effective bandwidth. node using FDDI bandwidth that been allocated node able achieve 100% available FDDI bandwidth. this section, have seen that when large buffer size used, sized according TTRT value ring, node will able achieve 100% efficiency with respect system network. throughput that node able achieve either limited bandwidth system when network load light, allocated FDDI bandwidth when network loaded, whichever smaller. FDDI adapter become essentially invisible between system FDDI ring.
Throughput TTRT Sized Buffers
following section examines throughput that might achieved example EISA node example FDDI ring when local buffer size used.
Unloaded Ring Throughput When FDDI ring unloaded, then node will achieve performance that virtually identical what achieved previously. When whole frame been transferred buffer, node starts trying capture token. Since tokens usable network access delay very small, node will capture almost immediately. Although system continued transfer data buffer while node waiting token, node waited for, most, During that time, much data transferred, relative maximum buffer size. buffer nearly full.
After capturing token, node transmits data buffer more that system supply during that time. Eventually buffer emptied node lets token system continues transfer data buffer again. token released early buffer memory never completely filled capacity. system bus, however, kept busy. constantly trying fill buffer that always being emptied network. Since system never stops
DOES SUPERNET STACK
SUPERNET second generation FDDI chipset available from Advanced Micro Devices, Inc. (AMD). SUPERNET provides built support buffer memory sizes ranging from hundreds bytes Expansion buffer memory easily accomplished because buffer memory implemented with external, standard, static RAM, available from number different memory vendors. This section shows that SUPERNET able support common methods buffer sizing, particular, sizing based TTRT. SUPERNET ideal chipset FDDI adapter card designs because allows designer tailor solution
Effect Local Buffer Memory Size FDDI Throughput Section 5.5, however, stated that after certain point TTRT dominant factor optimal buffer size. Rather, amount data that protocol will have outstanding time start dominate. Realistically, this amount data will order tens kilobytes. Given this, probably plenty.
meet appropriate market needs. Additionally, chipset easily provide software compatible solutions that span many different architectures target system loading environments. This results reduced development effort both hardware software which, turn, leads shorter time market family FDDI products.
CONCLUSION Latency Sizing
buffer used with SUPERNET half buffer allocated transmit data with other half reserved receive data, chipset able support maximum system latency This paper discussed some factors that must considered when sizing local transmit buffer memory FDDI adapter card. example based EISA FDDI adapter used demonstrate various concepts involved. shown that many simplistic methods sizing transmit buffer memory will cause great performance problems when FDDI ring becomes loaded. degradation caused network access delay, which increases dramatically more active nodes added FDDI ring. These methods buffer sizing take this effect into account. shown that sizing local buffer based negotiated Target Token Rotation Time allows node achieve maximum throughput available node will have throughput limited either system effective bandwidth allocated FDDI network bandwidth, whichever smaller. FDDI adapter will become invisible connection between FDDI ring host system. Finally, shown that AMD's SUPERNET chipset provides flexibility size local buffer meet many needs. Very important fact that SUPERNET easily allows designer meet buffer size requirements based protocol TTRT sizing. SUPERNET ideal solution family products that span number different operating environments price/performance characteristics. flexibility offered buffer sizing ability allows common core design become basis many products. core hardware design reused with each product software similarity provided chipset allows many designs completed quickly. This results great time market advantage.
(256 (1024 bytes/KB) ns/byte)
Most system busses have maximum latencies order hundreds microseconds, more than enough cover even most pathological operating conditions these architectures. Note that this calculation also neglects slight overhead claim beacon frames which permanently stored buffer memory.
Effective Bandwidth Sizing
When effective bandwidth used basis computing FDDI adapter buffer size objective simply insure that underrun maximum size FDDI packets prevented, given that effective system bandwidth less than 12.5 bytes/s. From buffer sizing formula that presented Section 5.2, know that size transmit buffer only needs 4500 bytes support correct operation. This size easily matched exceeded SUPERNET
TTRT Based Sizing
objective TTRT based local buffer sizing ensure maximum performance over wide range system FDDI ring loading environments. default maximum value TTRT FDDI data rate Mbps, this corresponds buffer size 2.06 bytes. Although this amount memory much larger than maximum buffer size supported SUPERNET realistic TTRT values will much lower than this. again, assume that half maximum buffer size SUPERNET used transmit buffer, then show, using equation Section 5.5, that SUPERNET support TTRT values 10.5 this case assumed that effective bandwidth system zero. assume effective bandwidth MB/s system bus, then buffer will support TTRT 15.8 while maintaining 100% efficiency.
References
Jain, "Performance Analysis FDDI Token Ring Networks: Effect Parameters Guidelines Setting TTRT," Proc. SIGCOMM'90 Symposium Communications Architectures Protocols, September 1990, Philadelphia, 264-275. EISA Specification, Version 3.0, BCPR Services, Inc., 1989 McCool, John "FDDI: Getting know inside ring", Data Communications, March, 1988. Primer FDDI: Fiber Distributed Data Interface, Digital Equipment Corporation, 1991.

David Roberts Publication Rev. Amendment Issue Date

Top Searches for this datasheet

Other recent searches