Performance Measurements in a Manufacturing Communication System

Communication System

Célio Vinicius Neves de Albuquerque

Marcelo Dias Nunes

Otto Carlos Muniz Bandeira Duarte

Universidade Federal do Rio de JaneiroCOPPE/EE - Programa de Engenharia Elétrica

P.O. Box 68504 - CEP 21945-970 - Rio de Janeiro - RJ - Brasil

FAX: +55 21 290.6626 - Email: otto@coe.ufrj.br

Abstract

This paper presents and analyses a low cost and high performance manufacturingcommunication system. It consists in the standard TOP profile implemented in singleprocessor computer connected to a IEEE 802.3 LAN. High throughput is achieved by anefficient implementation architecture based on specific layer interfaces and data structures,specialized mechanisms of memory management, timer management and task scheduling.Performance measurement results show a throughput efficiency that attains, for MAC Layeruser, 6 Mbit/s for a remote communication and 42 Mbit/s for loopback configuration. Athroughput of 2 Mbit/s is obtained for the Presentation Layer user. The most importantbottlenecks are analysed and consist, in important descending order, in the Transportacknowledgment step frequency, Transport checksum, LLC memory copy, memorymanagement and task scheduling.

1Introduction

Integration of the various elements in a manufacturing system has become an important issue. Thishas led to the development of Computer Integrated Manufacturing (CIM) concepts [1], whereproper communication is necessary to ensure adequate functioning of all the elements. CIM wasconceived in order to allow control and analysis of business data as well as technological data, fasterdesign and development cycles for more sophisticated products and flexible manufacturing systems.Two communication protocols, namely, Manufacturing Automation Protocol (MAP) [2] and

Technical and Office Protocols (TOP) [3], have emerged as a standard approach to enable easy andeffective communication between the elements in CIM environments. MAP/TOP specificationprovides a standard regarding all the aspects related to the communication between end systems.Rather than specifying new protocols, MAP/TOP are based on international standards adopted byISO, specifically the Reference Model for Open Systems Interconnection (RM-OSI) [4,5]. The RM-OSI defines a seven layered architecture structured in a hierarchical added value form.

The high bandwidth provided by the Local Area Networks (LAN) has moved the

performance bottleneck to the protocol processing time. Several works [6,7,8] have stated that highthroughput depends on a careful protocol implementation. Some authors exploit the intrinsicprotocol parallelism [9,10,11], others propose completely new protocols [12]. Nevertheless, thiswork deals with the problem of implementing the standard TOP communication profile efficiently ina low cost and single processor environment.

This paper presents an implementation architecture of a TOP profile and its performancemeasurements in a local area network (LAN) environment. In order to achieve high performancecommunication protocols, an implementation architecture is defined. This architecture is based onspecific layer interface and data structures, specialized modules of memory management, timermanagement and task scheduling.

The proposed implementation architecture privileges flexibility, modularity and weakcoupling between adjacent layers, minimizes the necessary retransmission buffers and takes intoaccount the memory fragmentation. The measurements show a good performance and analyse thethroughput due to each layer processing time, checksum mechanism and acknowledgment stepfrequency.

This work is organized as follows. Section 2 describes TOP profile and the implemented

features of each layer. Section 3 presents the implementation architecture. Its performancemeasurements are given in Section 4. Finally, conclusions are presented in Section 5.

2Manufacturing Protocol Profiles

CIM organizational hierarchy (Figure 1) is defined by the International Standards Organization (ISO)Reference Model for factory automation. This model supports six hierarchical levels: the EquipmentLevel (realization of commands to the shop floor equipment - robots, conveyors, vehicles, tools,etc.), the Station Level (numerical controllers, programmable logic controllers, etc., which direct andcoordinate the activity of the shop floor equipment), the Cell Level (supervision of the varioussupporting devices), the Section/Area Level (coordination of production and resource allocation),the Facility/Plant Level (production planning and scheduling), and the Enterprise Level (corporatemanagement).

OFFICE COMMUNICATION LEVELCAD/CAMWORKSTATION(S)TOPFACILITYHOSTCOMPUTERSTOPMAIN FRAMECOMPUTERSECTION / AREA LEVELMAPSECTION 1CONTROLLERSECTION 2CONTROLLERSECTION NCONTROLLERCELL LEVELMAPMICROCOMPUTERDATA CONTROL POINTMICROCOMPUTERDATA CONTROL POINTMICROCOMPUTERDATA CONTROL POINT# 1# 2# NSTATION LEVELMAPCNCCONTROLLERROBOTCONTROLLERPLCCONTROLLEREQUIPMENT LEVELROBOTFigure 1: CIM Organizational Hierarchy

MAP and TOP protocol profiles were specified as standards for manufacturing environments

allowing different shop floor and computer equipments to communicate. MAP first appeared in1986, when General Motors (GM) begun to apply the concepts of CIM to its operations, andrealized that communication was a bottleneck between devices on the plant floor. MAP has beenspecifically oriented to the needs of CIM, providing the communication which takes place on thefactory floor between programmable devices, cell controllers and area or section computers, asdepicted in Figure 1. The Boeing Company started work on the set of protocols which have evolved

ApplicationPresentationSessionTransportNetworkData LinkPhysicalMHSACSERTSEFTAMOSI Presentation ProtocolOSI Session ProtocolOSI Transport Class 4CLNSES-IS and X.25 PLPIEEE 802.2 (LLC)IEEE 802.3 (Ethernet)Figure 2: OSI Reference Model with the TOP profile

into TOP by the same time GM began defining MAP. TOP's main goal is to encompass everythingbut the manufacturing environment, accelerating the availability of multivendor economical,interoperable, off-the-shelf computing systems, devices and components. It supports the transfer offiles between different machines in the production and office environments, specifying standardizeddata formats. Figure 1 shows TOP's place in the organizational hierarchy and Figure 2, its layeredprotocol stack.

Below is a brief overview of each layer and the implemented protocol suite:

•

Medium Access Control (MAC) Sublayer: The Institute for Electrical and Electronic Engineers(IEEE) has created a family of lower layer protocols for LANs. This set of protocols is calledIEEE 802, and they are aimed at the needs of LANs. This work uses as TOP MAC Sublayer theIEEE 802.3 protocol - Carrier Sense, Multiple Access with Collision Detection (CSMA/CD) -for 10 Mbit/s Ethernet media. This sublayer directly accesses a set of freeware routines, knownas Packet Driver SPEC, which support a large number of communication controller cards.NE2000 drivers were used.

Logical Link Control (LLC) Sublayer: IEEE 802.2 standard specifies a logical link control foruse with any one of the other medium access control standards. There are three different typesof LLC. LLC Type 3 is acknowledged connectionless. LLC Type 2 is based on the High-levelData Link Control (HDLC) protocol. It requires a connection establishment and provides flowcontrol and error recovery functions. LLC Type 1 is unacknowledged connectionless and allows,besides point-to-point links, multicast and broadcast facilities. This implementation regards onlyLLC Type 1.

Network Layer: The Connectionless Mode Network Service (CLNS) is used with the LANspecifications. The reason for this choice is that the datagram approach is much more flexibleand robust when connecting multiple heterogeneous networks together, an important aspect ofTOP. The CCITT Recommendation X.25 Packet Level Protocol (PLP) was implemented,although it was not considered in this analysis. TOP also specifies the End System toIntermediate System (ES-IS) Exchange Protocol, which provides services performing dynamicrouting and updating of table information. However, this protocol will not be considered furtherhere.

Transport Layer: The purpose of a Transport Service in TOP environments is to provide reliableand optimized transparent data transfer between end nodes. This layer must do whatever isnecessary to bring the Quality of Service provided by the Network Layer up to the level requiredby the Transport Service users. OSI defines five classes of protocols (0 to 4) to guaranteetransferred data integrity, Class 4 being specified by TOP. Transport Class 4 provides aconnection oriented service with flow control, error detection (through checksum) and recovery(through retransmission), multiplexing/demultiplexing, segmentation/reassembly, numbering andconcatenation/separation facilities.

Session Layer: The standard for this layer is the ISO Basic Connection Oriented SessionProtocol. This protocol supplies user-oriented services of dialogue control,synchronization/resynchronization, activity management, negotiated release and half/full duplexdata transfer. TOP only requires implementations to provide the Kernel functional unit and theDuplex functional unit.

•

•

•

•

•

Presentation Layer: The ISO Presentation Protocol is used as a standard for this layer. Thepresentation layer allows many data representations to be used among communicating applicationentities. This adds much flexibility in system integration, while relieving the applications of dataformatting functions. Thus, it fulfills some requirements of CIM information formatting andsharing. Nevertheless, the considered presentation protocol implements only the Kernelfunctional unit.

Application Layer: This layer utilizes the services offered by lower layers to release theapplication process, so that the way in which data is addressed and delivered becomestransparent. It is within this layer that the system meets the end-user interface. Severalapplication service element standards were implemented: Association Control Service Elements(ACSE), Reliable Transfer Service Elements (RTSE), File Transfer Access and Management(FTAM) and Message Handling Service (MHS).

•

In this work, the throughput measurements were realized by a specialized user on top of eachlayer up to the Presentation Layer. Application Protocol performance measurements will beaddressed in future works.

3The Implementation Architecture

High performance and reliable data transfer in manufacturing environment are the main objectives ofTOP implementation. Since this system's first protocol implementation work [13], it became clearthat a careful implementation was necessary and that memory copy should be avoided. Hence anefficient implementation architecture was defined [14] in order to obtain high performance.

ApplicationASE1ASE - ASEInterfaceACSE - PRESInterfaceASE - PRESInterfaceASEASEN2ACSEMemoryManager

PresentationP - SInterfaceSessionTask

Scheduler

TransportT - NInterfaceS - TInterfaceTimerManager

NetworkN - LInterfaceLLCL - MInterfaceMAC

Figure 3: The Implementation Architecture.

3.1Interface Data Structures

There are essentially two basic ways of implementing interlayer communication: procedure calls andsynchronous/asynchronous message exchanges. Procedure call is certainly the most efficientmechanism but implies on a heavily coupled system. Synchronous interlayer communication is notsuitable for single process environments. Asynchronous communication is usually modeled by paired(input/output) message FIFO (first in, first out) queues.

This implementation architecture defines a standard interface between layers (Figure 3). Thisinterface consists of a pair of FIFO queues per connection and a set of data structure accessprocedures. It aims at normalizing the access to the services offered by adjacent layers, allowingflexibility, modularity and weak coupling between layers. This independence simplifies the divisionof the tasks of implementation, maintenance and updating of each layer protocol and leads to a bettersystem management.

Each FIFO element (Figure 4) is represented by a fixed size structure called PrimitiveStructure containing the following three fields:

•Primitive Structure Pointer - used for linking with the next FIFO element;•Primitive Name - name of the incoming/outgoing service primitive;

•Parameter Structure Pointer - pointer to a structure that contains the current primitiveparameters which is called Parameter Structure.

Primitive Structure

Primitive Structure PointerPrimitive NamePrimitive Parameter PointerParameter StructureParameter Structure PointerParameter PointerUser Data LengthUser Data PointerNumber of AlocationsUser DataParameterFigure 4: Interface Data Structures.

A connection setup corresponds to instantiating a pair of queues in the interface through

where all service primitives corresponding to that connection are transferred. However, ApplicationLayer is unique in the sense that it is composed by a sort of Common (ACSE, ROSE, RTSE) andSpecific (MHS, FTAM, ...) Application Service Elements (ASE) which can be associated in manyways with each other and with the Presentation Layer. The Application-Presentation Layercommunication is performed by ACSE and another service element. Therefore, some queues werecreated, namely, the ACSE-PRES and ASE(n)-PRES queues in the Application-Presentationinterface, and the ACSE-ASE(n) and ASE(n1)-ASE(n2) queues for the interface between ApplicationService Elements, where n identifies a service element.

3.2Protocol Control Information - PCI

A common processing for all layers, related to the communication between peer entities, is toinsert/remove PCI into/from SDUs (Service Data Units). There are many methods of inserting PCIto SDUs [8,14], for example, copy, maximum allocation and chaining.

SDUPCISDUSDUPCIPCISDUPCICopyPCISDU(a) copy (b) maximum allocation (c) chaining

Figure 5: PCI inserting techniques.

The insertion of PCI by copy (Figure 5a) consists in:

1.allocating a memory region with a length corresponding to PCI plus SDU sizes;2.filling up the PCI and copying the SDU;3.releasing the previous SDU.

Copying data to and from buffers between protocol layers in general represents significantoverhead [8,13,14,15]. Besides, the copying time is heavily dependent on SDU size. All these havea critical impact on performance.

The maximum allocation technique of inserting PCI (Figure 5b) consists in allocating a

sufficient memory region to hold the highest layer SDU plus the headers of all lower layers.Nevertheless, this procedure can lead to a great memory waste, because the maximum frame lengthwould necessarily have to be reserved. This scheme also introduces the need for copies to performsome basic mapping functions like segmentation and concatenation. Another argument against theuse of this procedure is the need for the upper layer to know the maximum frame size that can besent. This fact goes against the layer independence principle supported by OSI model [5,16].The chaining method of inserting PCI (Figure 5c) avoid data copy and waste of memory,consisting in:

1.allocating and filling up a PCI-size memory region;2.associating this region to a chaining structure;

3.chaining this structure to the received SDU structure.This approach seemed to be the best way of inserting PCI. Each queue is formed by a basic

element composed of chained data structures, defined in order to avoid data copy, a relevant factorof degradation in communication system performance. Figure 4 shows the implemented datastructures which base upon pointers and sets apart fixed size structures from variable size structures,to make possible a more efficient memory management. To implement this feature, the ParameterStructure was defined [14,19] containing the following five fields:

•Parameter Structure Pointer - used for linking with other Parameter Structure;•User Data Pointer - pointer to (part of) the primitive user data parameter;•User Data Length - length in bytes of the respective user data parameter;•Parameter Pointer - pointer to the other primitive parameters;•Number of Allocations.

The Number of Allocations field plays an important role for the system. It avoids several

retransmission copies when protocols based on error recovery by retransmission were used (DataLink, Transport, Session and RTSE protocols). This field is incremented/decremented every time alayer allocates/deallocates that structure. The corresponding retransmission buffer is released frommemory only when this field reaches zero. The Memory Manager is responsible for controlling thisproperty.

The interface between the MAC and LLC Data Link sublayers comprises a pair of circularbuffers, where the data to be sent to the communication controller are copied in a contiguous way.The controller is responsible for transmission over the physical medium.3.3

Memory Management

The insertion/removal of PCI into/from SDUs by each layer (or sublayer) necessarily requiresan efficient memory management, with the purpose of attaining a good performance. It is alsoimportant that the segmentation/reassembly of data be possible without the need for copies, but justwith some pointer manipulation. These constraints have led to the implementation of a specific andoptimized Memory Manager [19].

In order to avoid the excess of fragmentation, this module divides memory into transmission

and reception regions. It is possible because transmission and reception flows are independent.These regions are still subdivided in three parts (Primitive Structure, Parameter Structure and dataregions), allowing fixed and variable size memory allocations/deallocations. Optimized algorithmsare used, taking into account the specific behavior of an OSI system.3.4

Task Scheduling

The system's first version had a main procedure which checked whether the interface queues

were not empty, i.e., if there were any elements (service primitives) to be processed. This procedureproved to be too much time consuming. A great improvement in performance was achieved with thecreation of a Task Scheduler module [19]. This module consists of a single queue holding thesequence of tasks ready to be executed.

A protocol is easily defined by its State Machine, where from a certain event (receipt of aservice primitive) and the \"current state\some tasks are executed (or not) and another state isreached (or not), originating (or not) one or more new events (sending of primitives to adjacentlayers). This way, in the architecture described, every time a primitive is put into one of the interfacequeues (creation of an event), the corresponding execution task for this primitive is put into the TaskScheduler queue. The atomization required in the state machine execution is also easily obtained.3.5

Timer Management

Protocol implementation includes timer-based procedures which are used in error recovery

routines, connection establishment and release time-out control, etc. The Timer Managerimplementation follows Varghese and Lauck proposal [20], where an analogy with a real clock issimulated by a cyclic counter. This implementation proved to be very efficient, avoiding unnecessarylogical checks at each hardware interrupt, thus allowing the creation of a great number of run-time

timers. In the implemented architecture, a time-out is considered as a service primitive, that is, time-out indication (event) is put into the appropriate interface queue, and the corresponding task into theTask Scheduler queue.

4Performance Measurements

The implemented system is coded in C language and runs on IBM-PC-like personal computerenvironments using DOS operating system. Table 1 shows the number of code lines per layer and itsrespective object code size (without debug information). The basic measurement scenario consistedof two identical 33 MHz 486 DX processors connected to a IEEE 802.3 local area network throughNE2000 network adapters.

Table 1: Number of code lines per layer and its respective object code size.Layer / ModuleMemory ManagerTask SchedulerTimer ManagerMACLLCCLNSTP4SessionPresentationCode Lines Number9651553501120124022903560151302930Object Code Size(bytes)9 K2 K3 K3 K4 K15 K65 K177 K22 KThis performance analysis emphasizes on the throughput achieved by each layer user.Throughput is defined as the total useful information transferred (messages without headers andtails) per unit time. The total information amount transfer time should be measured from the instantthe originator issues the first message until the recipient receives the last one. However, as thisrequires a distributed measurement system, another measurement methodology was employed. Itconsists in computing, at the originator, the data transfer time interval between the first messagetransmission and receipt of the last message reply. The elapsed time measurement approach makesuse of PC tick-based function, providing a 55 ms accuracy. Therefore, in order to make themeasurement error negligible, an information amount of 10,000 frames is transferred. In case ofconnection-oriented protocols, no setup/release time measurements were performed. Furthermore,both computers were dedicated to the data transfer between them. The transfer is memory-to-memory to eliminate any effects from external I/O devices.

The measurements also include overheads introduced by lower layer headers. This overheadreaches 33 bytes for Presentation Layer user. In fact, the measurements are lower boundthroughputs because they include the user processing time due to message creation/destruction,message insertion/removal into/out of queues, memory allocation/deallocation, etc.

Figure 6 shows the throughput, in kbit/s, for the Packet Driver, MAC, LLC and CLNS Layer

users. Various message sizes are used, ranging from 128 bytes to 1024 bytes.

65Throughput (Mbit/s)432100

128

256

384

512

640

768

896

Size (bytes)

Packet DriverMACLLCCLNSTP4SessionPresentation1024

1152

Figure 6: Throughput versus message size.

The maximum theoretical IEEE 802.3 LAN throughput is 6Mbit/s and 9.8 Mbit/s for

minimum and maximum frame sizes, respectively. These throughputs may not be achieved by mostof the available adapters. A specialized user software was developed to measure the Packet Driveraccess routines performance. Its throughput is practically due to data transfer from microcomputermemory to controller card (through copy or direct memory access), and vice-versa, since thestandard interface, the memory and timer managers and the task scheduler were not utilized. By thisway, the time spent specifically with the transmission routine (Tsend) can be estimated. For example,with 1024 byte frame: Tsend = Tmem + Ttx = 1.4 ms, where Tmem is the time spent in data copyfrom memory to card, and Ttx is the time spent by the card to effectively send the data. In general,this is the initial system limitation.

SenderReceiverSenderReceiverPresentationPresentationPresentationSessionSessionSessionTranportTranportTranportCLNSCLNSCLNSLLCL - MInterfaceMACMemory Copy

(a)

LLCL - MInterfaceMACCommunication Medium(b)

LLCL - MInterfaceMACFigure 7: Measurement Scenarios. (a) Internal Loopback, (b) Connected by LAN.

As depicted in Figure 6, as the frame size increases, throughput reaches approximately6 Mbit/s. This result shows protocol processing time negligible when compared to transmission time(Tsend), for large frames, i. e., Packet Driver access routines and LAN adapter constrain systemthroughput. Hence, only single computer measurements will be considered further here. A specialloopback feature was developed at the Packet Driver level, consisting in frame copies from thetransmission buffer to the reception buffer, according to Figure 7. It is important to emphasize that,within this configuration, Tsend is replaced by the copy time Tcopy and transmission/receptionprocessing occurs on the same microcomputer. Just as an example, this time is about 0.22 ms for1024 byte frames.

Figure 8 shows that the system's performance improvement achieved with this newconfiguration. As an example, MAC sublayer throughput was about 43 Mbit/s, that is,approximately 7 times greater than the previous configuration throughput, limited by Tsend.

454035Throughput(Mbit/s)3025201510500

128

256

384

512

640

768

896

1024

1152

Size (bytes)

MACLLCCLNSFigure 8: Throughput versus message size (loopback scheme).

All communication controllers require sending data to be in a contiguous memory area. Thus,

LLC sublayer performs an unavoidable copy of chained data to a contiguous memory region(transmission buffer in LLC-MAC interface). This data copy is an important performancedegradation factor, responsible for a frame size dependent decrease in throughput.

Mem. Manag.

16%

Task Sched.

6%Others16%

Mem. Manag.

12%

Task Sched.

10%Others16%

Copy62%Copy62%

(a) originator (b) recipient

Figure 9: LLC sublayer processing a 1024 frame length.

In order to visualize the actual limitation introduced by copy, Figure 9 shows two pie charts

for the specific LLC sublayer processing. It is also observed that memory allocation/deallocation andTask Scheduler queue insert procedures spend a considerable processing-time portion.

Others35%

Others56%Mem. Manag.

32%Task Sched.

12%

Mem. Manag.53%

Task Sched.

12%

(a) originator (b) recipient

Figure 10: CLNS layer processing a 1024 packet length.

The CLNS Layer was initially implemented in an attempt to interact with connection-orientedtransport (TP4) and data link (LAP-D) services, with the purpose of mapping the transport/networkand network/data link interface primitives in CLNS primitives. This mapping requires many memoryallocations/deallocations, which consume a great portion of time, as can be depicted in Figure 10.

3TP4SessionPresentationThroughput(Mbit/s)2100

128

256

384

512

640

768

896

1024

1152

Size (bytes)

Figure 11: Transport, Session and Presentation user's throughput versus message size.

The Transport Layer is responsible for the greatest protocol performance decrease (Figure11). This was already expected because it is the Transport Layer in data transfer phase, for the TOPprotocol profile, that executes the greater number of functions [18]. Detection and error recoveryservices implementation requires the use of a checksum mechanism, and its computational overheadcan be quite high. Figure 12 shows the performance degradation introduced by the checksum. Thebar chart proves the strong dependence between checksumming and frame size.

Mem. Manag.

21%Task Sched.

6%

Others23%

Task Sched.

9%100Others33%8060Mem. Manag.

27%

4020Checksum50%Checksum31%

01282565121024 (a) originator (b) recipient (c) checksum vs. message size

Figure 12: Transport layer processing time for 128 bytes message length.

This protocol requires the transfer of Acknowledgment Transport Protocol Data Units (AK-TPDUs) for each DT-TPDU received, and this procedure overcharges system's processing.Acknowledgments are tied to window flow control, and a window size of 8 was used. Figure 13emphasizes the influence of acknowledgments exchange. To perform this measurements, checksumcomputing was disabled. Different acknowledgment steps were used, allowing one, two, three andfour outstanding TPDUs.

765

Throughput(Mbit/s)432100

128

256

384

512

640

768

896

1024

1152

Size (bytes)

TP4-AK4TP4-AK3TP4-AK2TP4-AK1TP4Figure 13: Transport user's throughput versus message size for various acknowledgment steps.Finally, the Session and Presentation Layers execute, in the data transfer phase, a commonprocessing for all layers: service primitives receipt, decoding and treatment. They spend a minimumprocessing time due to interface management (queue checks and elements insertion/removal),memory management (allocations/deallocations) and state machine execution, which cause aperformance loss at each layer.

5Conclusions

In this work an efficient TOP communication system was presented. Its implementation architecturewas described, which is based on:

•asynchronous message interlayer communications modeled by paired FIFO queues. This results

in a weak coupling between adjacent layers and a high modularity;

•interface data structures with pointers. Copies are avoided for interlayer communication,

segmentation and retransmission buffers;

•a memory manager that attains high performance separating transmission and receiving memory

regions. Fixed and variable size buffers allocations/deallocations are processed differently.Fragmentation is kept as a minimum;

•a task scheduler that avoids unnecessary queue tests and take advantage of required atomicity ofsome functions;

•a timer manager efficiently implemented and considered as a service primitive.

The measurement results show that the system performs well attaining almost the maximum

throghput at the MAC Layer user (6 Mbit/s). In a loopback configuration this layer throughput havereached 42 Mbit/s. Also in a loopback configuration the throughput for Presentation Layer user was2 Mbit/s, when TP4 was configured for an acknowledgment step of one and checksum computation.With an acknowledgment step of four and no checksum the Presentation Layer user throughputreaches approximately 6 Mbit/s.

The unavoidable copy realized by the LLC Layer has shown that memory bandwidth is an

important issue to obtain higher performane. It takes 62% of the total LLC processing time.The Transport checksum is too time consuming and justify a specialized hardware assists.

The time wasted in processing acknowledgments is very important. Blocking the acknowledgmentsinfluences the performance results. The throughput was multiplied by a factor greater than twowhen acknowledgment step frequency was divided by a factor of four.

The throughput results are quite good and fits well for most of the nowadays industrialapplications.

Acknowledgments

Prof. Marcelo Luiz Drumond Lanza, Prof. João Amaro Baptista Pereira, Luis Felipe Baginski, JoséFerreira de Rezende, Rainer Schatzmayr, Rogério Leone Teixeira da Cunha, Fernando MascarenhasCavalcanti de Barros, Frederico dos Santos Liporace, Calixto Damian Neto and Marcelo MacedoAchá have all contributed to the implementation of parts of the system.

References

[1][2][3][4][5][6][7][8][9][10][11][12]

L. J. McGuffin, L. O. Reid and S. R. Sparks, \"MAP/TOP in CIM Distributed Computing\IEEE Network, vol. 2, no. 3, pp. 23-31, May 1988.\"MAP 3.0 Implementation Release\".\"TOP 3.0 Implementation Release\".

The International Organization for Standardization, Data Processing, \"OSI SystemsInterconnection Basic Reference Model\

J. D. Day and H. Zimmerman, \"The OSI Reference Model\Proc. of the IEEE, vol. 71, no.12, pp. 1334-1340, Dec. 1983.

D. D. Clark, V. Jacobson, J. Romkey and H. Salwen, \"An Analysis of TCP ProcessingOverhead\IEEE Commun. Mag., vol. 27, no. 6, pp. 23-29, June 1989.

W. A. Doeringer, D. Dykeman, M. Kaiserwerth, B. Werner Meister, H. Rudin and R.Williamson, \"A Survey of Light-Weight Transport Protocols for High-Speed Networks\IEEE Trans. Commun., vol. 38, no. 11, pp. 2025-2039, Nov. 1990.

L. Svobodova, \"Implementing OSI Systems\IEEE J. Selected Areas Commun., vol. 7, no. 7,pp. 1115-1130, Sept. 1989.

O. G. Koufopavlou, A. N. Tantawy and M. Zitterbart, \"Analysis of TCP/IP for HighPerformance Parallel Implementations\Proc. 17th Conference on Local Computer Networks,Minneapolis, Minesota, Sept. 1992, pp. 576-585.

M. Zitterbart, \"Parallel Protocol Implementation for High Speed Networks\in Proc.SBT/IEEE Int. Telecommun. Symp., Rio de Janeiro, Brazil, Sept. 1990, pp. 260-264.

C. Papadopoulos, G. M. Parulkar, \"Experimental Evaluation of SunOS IPC and TCP/IPProtocol Implementation\IEEE ACM Trans. Networking, vol.1, no. 2, pp. 199-216, Apr.1993.

A. N. Netravali, W.D. Room and K. Sabnani, \"Design and Implementation of a High SpeedTransport Protocol\IEEE Trans. Commun., vol. 38, no. 11, pp. 2010-2024, Nov. 1990.

[13][14][15][16][17][18][19][20]

L. F. Baginski, F. M. C. de Barros, R. Schatzmayr and O. C. M. B. Duarte, \"Implementaçãoe Análise de Desempenho em um Sistema de Transferência de Dados\X Congresso daSociedade Brasileira de Computação, Vitória, Brazil, July 1991, pp. 157-169.

L. F. Baginski and O. C. M. B. Duarte, \"Um Modelo de Implementação de Alto Desempenhopara Sistemas Abertos\IX Congresso da Sociedade Brasileira de Telecomunicações, SãoPaulo, Brasil, Sept. 1991, pp. 1941-1945.

C. N. Woodside and J. R. Montealegre, \"The Effect of Buffering Strategies on ProtocolExecution Performance\IEEE Trans. Commun., vol. COM-37, pp. 545-554, June 1989.L. Svobodova, \"Measured Performance on Transport Service in LANs\Comput. NetworksISDN Syst., vol. 18, no. 1, pp. 31-45, 1989.

A. S. Tanenbaum, \"Computer Networks\W. T. Strayer, A. C. Weaver, \"Performace Measurements of Data Transfer Services inMAP\IEEE Network, vol. 2, no. 3, pp. 75-81, May 1988.

L. F. Baginski, \"Ambiente de Implementação para Sistemas de Alto Desempenho\MasterThesis, PEE-COPPE/U.F.R.J., Jan. 1992.

G. Varghese and T. Lauck, \"Hashed and Hierarchical Timing Wheels: Data Structures for theEfficient Implementation of a Timer Facility\in Proc. 11th ACM SIGOPS Symp. onOperating Systems Principles, Austin, TX, 1987, pp. 25-38.