The Protocol Bypass Concept for High Speed OSI
`Data Transfer
`
`C. M. Woodside*, K. Ravindran** and R. G. Franks*
`* Telecommunications Research Institute of Ontario, Dept. of Systems and Computer
`Engineering, Carleton University, Ottawa, Canada K1S 5B6
`** Dept. of Computer and Information Sciences, Kansas State University, Manhattan
`KS 66506
`
`Abstract
`A protocol bypass is a fast processing path which is used for some data units, for instance
`for large data packets. It can provide conformance with standardized layered protocol
`system like OSI, together with some of the performance benefits of a “lightweight” pro-
`tocol. The concept is discussed as it applies to the movement of user data by the OSI
`transport and session protocols, with some implementation experience, and an outline
`for an approach to formally deriving bypass specifications from protocol specifications is
`given. This outline uses some steps which must still be proven to be practical. Correct
`interleaving of data units from the two paths is a major concern, especially with mul-
`tiple asynchronously specified layers. It seems that the difficulty can be overcome and
`the concept has promise. In the (rather limited) implementation, bypassing consistently
`outperformed parallel processing as a means of performance enhancement.
`
`1 Introduction
`The search for generality, flexibility and standardization in communications protocols
`has led to the OSI layered system [1],[2] and many offshoots such as MAP [3]. However,
`the slowness of execution of the protocol implementations, which is essentially due to the
`complex checking of conditions that is done at every operation, is becoming a limiting
`factor in some applications. Therefore a new generation of lightweight or high-speed data
`transfer protocols is now emerging.
`The lightweight protocols exploit the low error rates of many networks, and use larger
`packets, reduced options to unclutter the data path, and more efficient methods for
`congestion control. They are partly motivated by the high speeds of new fiber-based
`networks, partly by high-throughput applications such as file system backups and full
`motion video, and partly by performance constraints already being felt with current
`protocols. Examples are Zwaenepoel’s Blast protocol [4, 5], VMTP [6], and XTP [7].
`Present-day “heavyweight” protocols such as OSI place notable performance con-
`straints on distributed applications. Svobodova [8] surveys the status of transport proto-
`cols (OSI and others) running on LANs and finds throughputs up to a maximum of about
`2 Mbits/s (other references are found in her paper). The goal of the current lightweight
`
`1
`
`Ex.1038.001
`
`DELL
`
`

`

`Complex cases
`(20% of cases?)
`
`Special simple
`cases (80%?)
`
`Usual
`Implementation
`
`Bypass
`
`Figure 1: The Bypass Concept
`
`protocols is to obtain an improvement by an order of magnitude now, and much greater
`increases later on faster processors and nets, into the range 100–1000 Mbits/s. The
`protocols do not have to be completely new ones; Clark et al [9] have analyzed the fun-
`damental limitations in IP/TCP to show that it could be operated in the 100 Mbit/s.
`range if implemented properly. They describe a “fast path” concept in [9],[10] which has
`already, in an existing workstation environment, moved files at close to the capacity of
`an Ethernet. Jacobson has called his guiding principle for a fast receiving path “Header
`Prediction”.
`The bypass concept described here is an attempt to generalize Jacobson’s “Header
`Prediction” to other protocols, particularly to layered protocol systems like OSI. Figure 1
`shows the general notion of The bypass part is activated when a very limited set of
`conditions apply which may be executed quickly; it will be effective if the conditions
`are satisfied often enough. Hopefully, one can obtain both the performance benefits of
`a lightweight protocol, when operating within a restricted set of circumstances, and the
`standardization and functional benefits of a fully conforming standard protocol stack.
`To obtain the full potential performance, some adaptation of the OSI standards may be
`necessary (e.g., larger window sizes). The bypass is an implementation concept rather
`than a new protocol, but it raises interesting questions about specification, such as
`• specifying the bypass path and control when several protocol layers are involved,
`• specification properties which could constrain the applicability of the concept in
`
`2
`
`Ex.1038.002
`
`DELL
`
`

`

`some cases,
`• formal derivation of the bypass process specification from the protocol specification.
`The bypass concept is based on a standard idea in performance enhancement, called
`the ‘centering principle’ by Smith [11] and sometimes called the “80–20 principle”. In
`‘centering’, special fast processing is provided for a frequently-occurring case. Although
`the idea is standard, and seems to be relatively straightforward in [11], it is nontrivial
`for layered protocols because of the sequential relationships generated by segmenting
`and reassembly of data, the complex interchanges supported by each layer, the web of
`relationships that may exist between entities at different layers due to multiplexing of
`connections, and especially because of the asynchronous nature of the separate layers.
`The purpose of this paper then is to explore the feasibility of constructing a bypass for
`an arbitrary layered protocol, and more particularly for OSI. It begins with an informal
`outline of a particular bypass for the OSI Session and Transport layers, identifying diffi-
`culties associated with operating and controlling a bypass, that seem to be fundamental.
`Ways to overcome these difficulties are discussed, and have been demonstrated by some
`preliminary implementation experience discussed in Section 6. The formal derivation
`of bypass specifications is approached in Section 7 by outlining how the simple control
`strategy used in Sections 3, 4 and 5 could be described formally. The paper is intended
`to show that bypassing is a fruitful and important mechanism for combining standard
`and lightweight protocol concepts, and that this topic is worth pursuing further.
`
`2 The Architecture of a Bypass
`Figure 2(a) illustrates the architecture of a bypass, including the following nomenclature:
`stack
`the set of entities implementing the standard protocol, which are to be bypassed,
`
`the user of the topmost layer of ‘stack’, assumed to be unique,
`user
`provider
`the service provider entity below the bottom layer of ‘stack’, also assumed
`to be unique at each end. These embody a service layer serving both ends, and
`hiding the physical transmission of data, etc.
`The Figure shows a system to support one-way data transfer, in which data flows from
`the ‘Data Sender’ end A to the ‘Data Receiver’ end B; note that control packets are also
`involved and flow in both directions.
`As a data unit flows from left to right, it is filtered by tests and if it qualifies, is
`processed by the fast path. The fast path is envisaged as a single process without
`internal concurrency, cutting through all the layers of ‘stack’, while ‘stack’ may have
`internal concurrent processes. Jacobson’s ‘Header Prediction’ conforms somewhat to
`this model, for in his system a received packet is examined for a match against a stored
`predicted header at the test labelled RBYPASSTEST, but there is no SEND test.
`One of the potential values of isolating the fast path rather than just optimizing the
`code in ‘stack’ is the excellent prospect of implementing the bypass in hardware. Fig-
`ure 2(b) shows a suggested division when the service entities are hardware-implemented
`link controllers such as Ethernet controller chips. SBP and RBP are the ‘send’ and ‘re-
`ceive’ bypass devices. The receive test RBYPASSTEST is in hardware, to avoid breaking
`the hardware path to the user level, and the RBP device writes data into addresses sup-
`plied in memory accessed by the device. The send test does not have such compelling
`reasons to be in hardware, and so is not.
`
`3
`
`Ex.1038.003
`
`DELL
`
`

`

`(a) Architecture of Bypass for Sender and Receiver
`
`user A
`
`user B
`
`SPS
`
`SPS
`
`Link
`
`Net
`
`Link
`
`shared
`data
`
`Send
`Bypass
`Device
`
`shared
`data
`
`Receive
`Bypass
`Device
`
`(b) A Hardware Implementation Strategy with two bypass devices
`(for the case where the provider is the link controller)
`
`

`

`3 One-Way Bypass of OSI Session and Transport
`Layers
`To focus the discussion this section examines a bypass around an OSI session layer
`combined with a Transport Class 2 layer, and used with extended TPDU-numbering,
`but without flow control.
`
`3.1 Ideal Data Transfer
`Further, bypassing will be provided only for those data units and transfers which satisfy
`the following conditions for “ideal data transfer”,
`
`• the session service data units are of type “data”,
`• there is no segmentation/reassembly, concatenation/separation, or multiplexing in
`the two layers,
`• there are no piggybacked acknowledgements or window credits
`
`It
`Attention is also restricted to a one-way bypass, as illustrated in Figure 2(a).
`could be used for instance in a file server with a send-only bypass loading pages across a
`network to diskless workstations with receive-only bypasses. Only one-way “fast paths”
`were considered in [9, 10]. Greater generality will be introduced in later sections.
`In an “ideal data transfer” each SSDU corresponds to exactly one NSDU and entails
`these operations in the layers:
`
`• increment PDU numbers by one in each layer at sender and receiver,
`• send an acknowledgement, after a receive.
`The bypass must produce the same effects as a standard implementation.
`
`The general architecture of a one-way bypass system is as shown in Figure 2(a).
`
`3.2 The Bypass Tests, without Window Flow Control
`Correct execution of the protocol by a bypass depends on two key operations – the iden-
`tification, at the bypass-condition tests in Figure 2(a), of those data units which which
`will give an ideal transfer, and correct interleaving of data units which follow the two
`paths. On sending a user data unit, the conditions to be met are that it be small enough
`to fit into one TPDU, and that it be a ‘data’ type of unit (rather than control). The
`connection must also meet the conditions stated above (eg., no multiplexing). Therefore
`the bypass test will be only partly based on headers (as it was in Jacobson’s method); it
`will also include some aspects of the state of the protocol layers, shown as the “shared
`data” in Figure 2(a).
`Packets which fail the bypass test follow the normal processing path through the full
`implementation. This raises the problem of ensuring correct interleaving of data units
`from the two paths where they rejoin. This is a non-trivial problem because in principle
`the separate layers in OSI are asynchronous entities. In this work interleaving is ensured
`by further assuming or enforcing that the protocol execution for an ideal data transfer
`is atomic with respect to all other operations (send or receive, and by either layer) for
`the same connection. This extra assumption is called ‘relative atomicity’ of processing.
`Then there is a simple toggle in the flow, switching from one path to the other, and the
`
`5
`
`Ex.1038.005
`
`DELL
`
`

`

`initiation of the new path at each switch is delayed until the former path is empty. This
`delay is simple to control when switching away from the bypass because it is a relatively
`simple process; the delay must be carefully enforced for switching the other way.
`To be effective, the bypass must go around every layer with an important processing
`load, and the window size must be large enough to make the extra speed useful. Larger
`window parameter values than those now standard in OSI would be needed to fully
`exploit the concept. Bypassing of link processing would assume a capability for large
`link-level data units, to avoid segmenting at the link level. In [9], for example, link data
`units of 8K bits or more were used.
`Initially, the tests will be given for a case without window flow control; it will be
`added later in this section.
`
`Sender
`Besides the usual protocol entities for Session and Transport layers, there is a send bypass
`entity SB, as shown in Figure 2(a), which only operates when sending a data unit from
`the user. The bypass entity will be associated with the session service access point,
`and will communicate directly with the network (provider) SAP for the connection that
`is being bypassed. A particular set of entities (session, transport and bypass) will be
`associated with a given user-to-user connection by an initiation mechanism considered
`below.
`The following protocol entity state data must be shared with the bypass entity (re-
`member that we are ignoring window operation at this point):
`
`• session layer major state (read-only by bypass process)
`• SPDU-NR, next session PDU number for sending (read-write by bypass process)
`• session connection identifier (read-only by bypass process)
`• transport layer major state (read-only by bypass)
`• TPDU-NR, next transport PDU number for sending (read-write by bypass)
`• max TPDU size (read-only by bypass)
`• transport connection identifier (read-only by bypass)
`• network entity information for sending (eg. SAP identifier) (read-only by bypass)
`
`Besides the above, two test conditions must be accessible to the bypass test:
`
`• session entity empty (entity is idle and no data units are ready for processing in
`either direction)
`• transport entity empty (entity is idle and no data units are ready for processing in
`either direction)
`
`Then the bypass test and operation for send has the form:
`
`SBYPASSTEST = (session layer state = DT) and
`(session entity empty) and
`(transport layer state = DT) and
`(transport entity empty) and
`(predicted TPDU length <= TPDUmax)
`
`if SBYPASSTEST then addheader; send NSDU; update; else send SSDU;
`
`6
`
`Ex.1038.006
`
`DELL
`
`

`

`Session Header
`
`Transport Header
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10
`
`DST|CDT
`
`DST-REF
`
`TPDU-NR and EDT
`
`15 | 0
`
`ref
`
`counter, and EDT = 0
`
`3
`
`LI
`
`7
`
`2 L
`
`I 0
`
`1 S
`
`I 1
`
`octet
`
`field name
`
`value
`
`fixed values
`for the one connection
`
`variable values (counter)
`
`Figure 3: Session and Transport Header Format for the Case of Section 3
`(class 2 transport with extended TPDU-NR and EDT format, no segmenting/reassembly
`so EDT=0, no flow control so CDT=0)
`
`where “send NSDU” sends to the network entity, while “send SSDU” sends to the session
`entity (for non-bypassed units), and “update” increments SPDU-NR and TPDU-NR. The
`operation “addheader” adds the combined headers for the two layers to the data in the
`form shown in Figure 3, to give exactly the same TPDU that would be constructed by
`the two layers together. In the restricted operational cicumstances of this bypass a fixed
`header template, with only one variable field for TPDU-NR, can be retained for fast
`processing. Acknowledgements may be handled by SPS.
`
`Receiver
`At the receiver, the following protocol entity state data must be shared:
`
`• session user entity information for delivering SSDUs
`• session layer state (read-only by bypass process)
`• SPDU-NR next session PDU number for receiving (read-write by bypass process)
`• session connection identifier (read-only by bypass process)
`• transport layer state (read-only by bypass)
`• TPDU-NR next transport PDU number for receiving (read-write by bypass)
`• last TPDU acked (read-write by bypass)
`• transport connection identifier (read-only by bypass)
`• network layer information for receiving (read-only by bypass)
`
`The same two layer state test conditions as before must be accessible to the bypass
`test:
`
`• session entity empty
`• transport entity empty
`
`The bypass test is executed on receipt of an incoming NSDU from the network:
`
`7
`
`Ex.1038.007
`
`DELL
`
`

`

`RBYPASSTEST = (session layer state = DT) and
`(session entity empty) and
`(transport layer state = DT) and
`(transport entity empty) and
`(header match)
`
`The header match condition ensures that the data is for the right connection, and has
`the expected sequence numbers. It is based on a predicted and stored representation, as
`illustrated in Figure 3. The bypass operation for receive has the form:
`
`if RBYPASSTEST then stripheader; deliver SSDU; update; else deliver NSDU;
`
`where “deliver NSDU” sends to the transport entity, while “deliver SSDU” sends to the
`session user entity. The operation “stripheader” removes the header component from the
`data as a standard, preprogrammed number of octets, and if necessary does the transport
`level cyclic redundancy check. The operation “update” increments TPDU-NR, SPDU-
`NR and issues acknowledgements as NSDUs (using a preconstructed format into which
`the PDU numbers are inserted). Each time an AK is issued the “last TPDU acked” value
`is updated.
`
`Comments on Operation of a One-Way Bypass
`The specification above assumes that the bypass entity shares data with the standard en-
`tities, and that process scheduling can enforce the relative atomicity of bypass processing
`with respect to the standard entities it is associated with.
`Acknowledgements are received by the standard entities, and their handling is not
`accelerated. Since every packet is acked this may become a significant load, and the traffic
`through the standard entities may interfere (through the “empty entity” conditions) with
`availability of the sending-side bypass. Ideally, by the appropriate use of priorities or by
`some other mechanism, the receipt of the AK by the standard entities should always be
`completed before the test for the next send is executed. Then a long sequence of user
`data packets which meet the conditions on length should all pass through the bypass.
`Both the above points indicate how the specification of a bypass interacts more closely
`with implementation considerations than is the case with the OSI standards. This is to
`be expected when the existence of the bypass is governed by performance goals. The
`intention here however is to specify the implementation considerations only as formal
`constraints, without saying how they are to be satisfied.
`
`3.3 Additional Features
`Adding Window Management
`We can bypass with window flow control by adding
`1. to the shared data at the sender:
`• available transport window credit (read-write by bypass process)
`2. to the SBYPASSTEST
`• the condition (window credit available)
`
`8
`
`Ex.1038.008
`
`DELL
`
`

`

`The actual window management can still be done by the transport entity when it
`handles an AK, or other incoming PDU with window credit attached. The sender bypass
`rejects the unit if there is no window credit, and the data is queued by the transport
`entity to be sent by it on receipt of credit. Until the backlog is cleared the (transport
`entity empty) condition will fail.
`This description provides window functionality, but if the window ever closes the
`slower path will be used and the sender may never catch up with the backlog, and may
`never be able to re-open the bypass. To avoid this problem the bypass would itself have
`to handle AKs and window credits, as discussed under the two-way version below.
`
`Initiation of Bypass Flow
`Bypass operation is most useful for certain situations, such as bulk data transfers. At
`other times the test overhead may make it unprofitable. The appropriate conditions for
`initiating it may be detected and operated on in various ways, such as:
`• user initation, by an action from the application level;
`• from the quality of service parameters (QOS), as interpreted by the transport layer
`
`• by an automated process based on observing the packet stream.
`The details of the initiation process are not a major concern of this paper; any of the
`above approaches could be taken. They have different implications for affecting the
`implementation of the standard stack, however.
`
`4 Two-way Bypass
`Here we consider doing the bypass in both directions. The above specification of send
`and receive bypasses are combined into a single process (still atomic, but now triggered
`by events either at the session service user or the network service provider) so that each
`end has both the send and receive features. Then data transfer will be bypassed in either
`direction while AK and window operations will be carried out by the standard entities.
`We will now examine some details of the changes that would be required.
`
`Full AK Handling
`Further performance improvement might result, in cases with window flow control and
`error control, if AKs and window management could be handled by the bypass. For AKs,
`the receive bypass makes a second test if RBYPASSTEST fails, as follows:
`
`RAKTEST = (session layer state = DT) and
`(session entity empty) and
`(transport layer state = DT) and
`(transport entity empty) and
`(header match)
`(packet matches AK template)
`
`which leads to an alternative operation when RBYPASSTEST fails, as:
`
`if RAKTEST then ack-update;
`
`9
`
`Ex.1038.009
`
`DELL
`
`

`

`Here “ack-update” updates the “last TPDU acked” to the value in the AK, and up-
`dates flow control window credits as conveyed by the AK. Similarly, on RBYPASSTEST
`the “update” operation ends by processing piggybacked credits in the same way.
`On receipt of window credits in cases where there is a backlog of transport data units,
`the standard transport entity must then be forced to process its waiting data, as it would
`do on receipt of window credits in an AK.
`
`Window Management
`If on the other hand the bypass attempts to also send the waiting data units as part of
`a single ‘relatively atomic’ operation, it produces messy scenarios. On receipt of window
`credits the queue of waiting data may include units that cannot be bypassed, mixed with
`some that can. To sort out the bypassable units, a different bypass test is needed because
`the data has already been partly processed. Further, when one of the waiting units is sent
`through the standard entity it breaks up the atomicity of the bypass processing which
`was originally initiated by the receipt of the data unit with the window credits.
`A simple-minded solution is to block all data units at the bypass entrance when
`window credit is exhausted, and then when credit is received to force their processing.
`This forced processing would activate the bypass test on each of the waiting data units in
`turn until either the credit or the queue of waiting data is finished. Data units taking the
`standard path would be processed to completion before the next was tested, as discussed
`earlier, to permit the next one to be bypassed.
`
`5 Experience
`A bypass for the OSI Session and Transport Layers along the general lines described above
`has been programmed into a protocol performance testbed described elsewhere [12]. The
`standard entities in the testbed were implemented on parallel processors, and this both
`biased the test against the bypass and also complicated the sharing of data and the
`entity empty tests, which were implemented with inter-process messages. Nonetheless,
`even in this rather hostile framework, the bypass roughly doubled the throughput of a
`long sequence of data packets.
`The task architecture of the implementation is shown in Figure 4. A real-time kernel
`called Unison [13] was used; Unison has parallel tasks with send-receive-reply messages
`and a task can send to or receive from specific ports, shown as little parallelograms within
`the larger shapes representing tasks. The processors were MC68020’s on a VME bus.
`The two ends of the connection were run on the same backplane, so the net work delay
`is insignificant and flow control is not a problem; in any case only Transport level 0 was
`implemented so flow control was not an issue.
`The protocol bypass concept was applied to an implementation of the OSI session
`and transport layers written in C and running on a common bus multiprocessor. The
`original software consisted of four layers: a user aplication, a session and transport layer,
`and a vestigial network layer. (The latter was simply used to provide the proper interface
`to the transport layer and to connect the source and destination together.) This design,
`including the modification for bypassing, is illustrated in Figure 4.
`The protocol software must be capable of handling traffic in both directions simulta-
`neously, therefore layers cannot send messages directly to one and another. If they did,
`
`10
`
`Ex.1038.010
`
`DELL
`
`

`

`deadlock would arise when two layers tried to communicate with each other at the same
`time. Consequently, each layer consists of up to three tasks. The first task implements
`the actual functionality of the layer while the others are used to transport data from one
`layer to the next.
`For protocol bypassing, the original design was augmented by changing the network
`and application layer software to reroute messages directly to each other. The bypass is
`designed so that the user layer always attempts to send data using the bypass route, the
`exception occuring when data must be segmented. Incoming packets are initially routed
`through the main protcol software. However, once a data connection is established, the
`network layer, through the bypass software, also attempts to route data through the
`bypass channel. Packets which fail to match the bypassing criteria are routed in the
`normal manner. The modifications to the architecture are indicated by shaded portions
`in Figure 4.
`
`Experimental Configurations
`In the original design, up to eight processors were used: one processor was assigned to
`each of the session and transport entities, to each half of the user-level software, and
`to each half of the interface between protocol stacks. This configuration exploited the
`maximum degree of parallelism on the hardware. The protocol software could also be
`partitioned to run on only two processors.
`In this configuration, all of the software
`making up the source half of the connection ran on one processor, while the destination
`half ran on the other.
`The bypassing variant of this software was also designed to run on multiple processors.
`Since the bulk of the protocol bypass software is found in the user and network layers, no
`additional processors are needed. Furthermore, when the bypassing is operational, only
`four of the eight processors are active because the session and transport layer tasks are
`not used.
`
`Results
`Tables 1 and 2 present the throughput of the system when configured with one processor
`per party (two processors in the testbed) or four processors per party, (eight processors
`in the testbed) respectively. In the latter case each major task in Figure 4 had its own
`processor. These tests were conducted using a Class 0 transport connection sending data
`in one direction only. The performance improvement by using bypass ranges from 120%
`for small packet sizes to 75% for larger packet sizes, when using just two processors. When
`tested using the eight-processor configuration, the throughput improvement ranges from
`50% down to 25%.
`It is notable that a bypass using only one processor per party was uniformly better
`than a parallel implementation on four processors per party without bypassing. This is
`due to the inherent efficiency of the bypass concept. For messages of 500 bytes and over
`the multiprocessor implementation with the bypass performed worse than the unipro-
`cessor implementation. This appears to be because of bus saturation, since the bypass
`coordination and inter-layer data-passing both placed loads on the bus in the multipro-
`cessor case.
`
`11
`
`Ex.1038.011
`
`DELL
`
`

`

`Figure 4: Protocol Software Task Architecture (one end) The parallelograms represent
`tasks; the arrows are for the direction of sending messages. The shaded components
`implement the bypass (sending on the right, receiving on the left).
`
`12
`
`Ex.1038.012
`
`DELL
`
`

`

`

`

`SP S(u, l)
`
`user(u)
`provider(l)
`ES
`
`the set of protocol entities to be bypassed, defined as a process
`communicating through an upper gate u and a lower gate l,
`the user of the topmost service provided by SP S,
`the service provider to the bottom-most entity in SP S,
`the empty-stack condition, satisfied by SP S when all its
`protocol entities are blocked awaiting events and all
`its FIFOs are empty.
`We consider ES to be applied to operations and data units for a single connection.
`When ES is true and an event arrives at either u or l it initiates processing by SP S,
`and the period until ES is true again is a busy period for SP S.
`By adding some mechanism, such as a process scheduler or a screening process which
`filters the inputs to SP S, a restricted process SP S ′ (u, l) can be derived which always
`processes each input event at u or l until ES is true, that is until all possible interactions
`(internal and external) are exhausted, before accepting input again. Since protocol spec-
`ifications make no assumption about the arrival time of the next event, this restriction
`can always be contructed in such a way to still satisfy the protocol specification and
`without introducing deadlocks. That is, SP S ′ is a correct implementation of SP S. We
`will not attempt to define SP S ′ formally here. If we denote a generic external event as
`‘ul’ (meaning u or l) then one event is processed by ST ACK, defined as follows:
`
`This can also be expressed as a combination of a send process SP SS ′ for u, and an
`alternative receive process SP SR′ for l:
`
`ST ACK = ul → SP S ′
`
`STACK = ul → SP S ′ = u → SP SS ′|l → SP SR′.
`
`We will define the above notation to permit shared data between the alternative
`processes SP SS ′, SP SR′. Now we divide the event stream into components ub, lb and
`uln:
`ub for those ‘send’ events u for which the data associated with the events satisfies bypass
`conditions for sending (e.g., a small enough data unit),
`lb similarly for bypassable events l, for receiving,
`uln for all the other events, that cannot be bypassed,
`and we also introduce a pair of guard conditions gs and gr, which are satisfied if the
`internal state of SP S ′ is satisfactory for send bypass, and receive bypass processing,
`respectively. The process notation
`
`ub(gs) → SP SS ′
`
`represents an event ub and satisfaction of the guard condition gs together leading to
`SP SS ′. gs and gr must satisfy
`
`ub(gs) → SP SS ′ = ub → SP S ′
`lb(gr) → SP SR′ = lb → SP S ′
`
`Then
`
`ul → SP S ′ = ub → SP S ′|lb → SP S ′|uln → SP S ′
`ul → SP S ′ = ub(gs) → SP SS ′|lb(gr) → SP SR′|uln → SP S ′.
`
`(1)
`
`14
`
`Ex.1038.014
`
`DELL
`
`

`

`Finally we define bypass processes BS for send, BR for receive, in any way that satisfies
`
`ub(gs) → BS = ub(gs) → SP SS ′
`lb(gr) → BR = lb(gr) → SP SR′
`
`(2)
`
`The construction (2) implies that SP SS ′, SP SR′, BS and BR have the same restriction
`on concurrency as SP S ′. Combining (1) and (2) we obtain
`
`STACK = ul → SP S ′ = ub(gs) → BS|lb(gr) → BR|uln → SP S ′
`
`With recursion the resulting process X is
`
`X = (ub(gs) → BS|lb(gr) → BR|uln → SP S ′); X
`
`This discussion which ends here, opens up a number of further questions about formal
`descriptions:
`• how to formally describe all the necessary process properties (i.e., is LOTOS able
`to do it?);
`• how to specify SP S ′, SP SS ′, SP SR′;
`• how to specify the guard consitions gs, gr;
`• how to split the event stream into the components ub, sb, etc.
`• how to derive BS and BR from SP S ′.
`
`7 Effects of a Bypass on the Implementation of the
`Standard Stack SPS
`An attractive notion is to be able to add a bypass to any standard stack implementation
`without modifying the latter at all. This is possible in principle because the tests are
`external to the stack; the only requirement is sharing of the stack state data. Because
`mutual exclusion is provided by the relative atomicity and empty stack conditions, only
`the location and coding of the necessary state must be known. Because our own imple-
`mentation in Section 6 did data-sharing and established the empty stack condition with
`messages, it required extensive changes to the standard entitites.
`In practice it is somewhat doubtful that the empty stack condition can be established
`without at least a thorough knowledge of the implementation, as it is not part of the
`protocol standard.
`
`8 Conclusions
`The present paper has described a bypass or “fast path” concept which may be used
`with layered protocols. It has described it informally, described implementation expe-
`rience, and discussed how bypasses may be specified formally.
`It seems that, as an
`implementation technique, bypassing offers significant immediate performance payoffs.
`The specification problem raises some challenging issues, listed above.
`An interesting question opened here is that of deriving a projection of a protocol or
`other asynchronous process, into a “subspace” of processes constrained by more-or-less
`arbitrary conditions such as those for bypassing. Also a better technique for connecting
`and controlling the bypass may be found, which would not require the ES condition, but
`rather would allow full concurrency on at least one path.
`
`15
`
`Ex.1038.015
`
`DELL
`
`

`

`Acknowledgements
`Discussions with Luigi Logrippo and Moshe Krieger were helpful in f