|
This Technical Note explains how to effectively use timers and retry mechanisms
of the various AppleTalk protocols to achieve maximum performance on an
internet.
[Apr 01 1990]
|
Introduction
The most fundamental service in an AppleTalk internet is the Datagram Delivery
Protocol (DDP), which provides a best-effort, connectionless, packet delivery
system. A sequence of packets sent using DDP on an AppleTalk internet between
a pair of machines may traverse a single high-speed Ethernet network or it may
wind across multiple intermediate data links such as LocalTalk, TokenRing,
etc., which are connected by routers. Some packet loss is always inevitable
because of the loosely coupled nature of the underlying networks. Even on a
single high-speed Ethernet network, packets can be lost due to collisions or a
busy destination node. The AppleTalk Transaction Protocol (ATP), the AppleTalk
Data Stream Protocol (ADSP), and other high-level protocols protect against
packet loss and ensure reliability by using positive acknowledgement with
packet retransmission mechanism.
The basic transaction process in ATP consists of a client in a requesting node
sending a Transaction Request (TReq) packet to a client in a responding node.
The client in the responding node is expected to service the request and
generate a series of Transaction Response (TResp) packets, which also serves as
an acknowledgement. The ATP process in the requesting node also starts a timer
when it sends a packet and retransmits a packet if the timer expires before a
complete response arrives. In a large internet with multiple gateways, it is
impossible to know how quickly acknowledgements may return to the requestor.
If you set the retry time to be too small, you may be retransmitting a request
while a delayed response is en route, but if you wait too long to
retransmit a request, application performance may suffer. More importantly,
the delay at each gateway depends upon the traffic, so the time required to
transmit a packet and receive an acknowledgement varies from one instant to
another. To further complicate matters, two packets sent back to back could
take completely different routes to the destination.
Back to top
Selecting ATP Retry Time and Retry Count
You can use the round trip time for a transaction as a heuristic for setting
the retry time and retry count. The round trip time between two nodes in a
particular internet at a particular time is usually deterministic.
The easiest way to set the retry time is to assign a static value based on the
round trip time for a transaction. The AppleTalk Echo Protocol (AEP) can be
used to obtain the round trip time in a given moment between two nodes. AEP is
implemented in each node as a DDP client residing on statically assigned socket
number four. You should use DDP to send AEP requests through any socket that
is available, and you should use the maximum packet size that you plan on using
in your application. You can listen for AEP responses by implementing a socket
listener. The following code is an example AEP socket listener.
;_________________________________________________________________________
;_________________________________________________________________________
;
; EchoDude
;
; 3/90 pvh - MacDTS
;
; (c)1990 Apple Computer, Inc.
;
; The following MPW Asm code is a socket listener for reading in returned Echo
; (DDP type 4) packets.
;
; The target device was shipped a packet with a '1' in the first byte of the data
; area by way of a DDPWrite. It was sent to socket 4, the Echoer socket. If the
; target device has an Echoer, it will send a return packet to us of equal size
; except it will have replaced the '1' in the first byte with the value '2'. This
; indicates an EchoReply packet.
;
; The listener itself (RcvEcho) is added with a POpenSkt (Inside Mac V-513) call by
; passing the address of the listener in the listener field of the parameter block.
;
; All we really are trying to accomplish here is to set up a notification for
; returned packets from the target Echoer. A time (Ticks) is stuffed into a
; location our app can find (actually back into the packet buffer) and will be used
; to calculate round trips times. We'll also save off the hop count from the packet
; header for fun too since I have nothing better to do with my time on weekends.
;
; More could be done with this listener as far as making sure that we are only
; receiving back a packet from the node we sent it to etc.... but we can't
; encompass everything in a sample. Okay, well we could... but we have to leave
; something for you guys to do.
;
; It should be noted that careful preservation of register A5 is necessary.
; LAP requires that A5 be preserved AFTER the call to ReadRest. i.e. you
; cannot save A5 onto the stack when your socket listener is entered, call ReadRest
; and then restore A5 from the stack and exit. Wah. LAP requires that the address
; placed in A5 during ReadRest be there when your socket listener is exited.
; So... if you need a different A5 after the call to ReadRest make sure you restore
; it before RTS-ing back the caller.
;
;
; Called:
; A0,A1,D1 : Preserve until after ReadRest
; A2 -> MPP local variables
; A3 -> RHA after DDP header
; A4 -> ReadPacket, 2(A4) -> ReadRest
; A5 Useable until ReadRest
; A6,D4-D7 : Preserve across call
;
;__________________________________________________________________________
EchoSkt EQU 4 ; Echo socket number
EP EQU 4 ; EP DDP protocol type
EPReq EQU 1 ; Code for echo request
EPReply EQU 2 ; Code for echo reply
;
; Read the packet into the echo buffer
;
RcvEcho PROC EXPORT
EXPORT our_A5 : CODE
EXPORT our_Buff : CODE
IMPORT GBOB:DATA
BRA.S checkEcho
our_A5
DC.L 0
our_Buff
DC.L 0
our_Hops
DC.W 0
our_Ticks
DC.L 0
checkEcho
CMP.B #EP,-(A3) ; Make sure it's an echo packet
BNE.S RcvEIgnore ; Ignore it if not
LEA toRHA(A2), A3 ; top of RHA
CLR.L D2 ; clean up D2
MOVE.B lapType(A3), D2 ; lap type
CMP.B #longDDP, D2 ; check for long header (Type #2 packet)
BNE.S noHops ; wah... no hops if short packet
MOVE.B lapType+1(A3), D2 ; this is the hop count byte, 1st byte in DDP
; header
AND.B #$3C, D2 ; mask to middle 4 bits of byte for hop count
; | x | x | H | O | P | S | x | x |
ASR.B #2, D2 ; shift 2 bits to right
LEA our_Hops, A3 ; address of our storage
MOVE.B D2, (A3) ; move # of hops into our storage
noHops
MOVE.W #DDPMaxData, D3 ; our buffer is #DDPMaxData in size
LEA our_Buff, A3 ; address of buffer to read packet into
MOVE.L (A3), A3 ; set buffer
JSR 2(A4) ; ReadRest of packet into buffer
BEQ.S RcvEchoReply ; If no error, continue
BRA.S RcvEchoFail ; dang...
RcvEIgnore
CLR D3 ; Set to ignore packet
JMP 2(A4) ; Ignore it, ReadRest and return
BRA.S RcvEchoFail
RcvEchoReply
CMP.B #EPReply, -DDPMaxData(A3) ; make sure it's our reply packet
; it shouldn't be anything else, but check
; just in case
BNE.S RcvEchoFail ; if not our reply then blow
MOVE.L A5, D2 ; save dude in D2
LEA our_A5, A5 ; address of our A5 local storage
MOVE.L (A5), A5 ; make A5 our A5 for application global use
MOVE.B #1, GBOB(A5) ; set flag confirming reception of
; echo reply packet
LEA our_Buff, A3 ; address of our local buffer storage into A3
MOVE.L (A3), A3 ; get saved pointer and set buffer.
LEA our_Hops, A5 ; address of hops local storage... notice we
; are TRASHING A5 with this!!!!!
MOVE.W (A5), (A3)+ ; copy in hop count to buffer
MOVE.L Ticks, (A3) ; next copy in Ticks
MOVE.L D2, A5 ; restore dude
RTS ; return to caller
RcvEchoFail
RTS ; return to caller
ENDP
setUpSktListener PROC EXPORT
IMPORT our_A5 : CODE
IMPORT our_Buff : CODE
LEA our_A5, A0 ; this copies
MOVE.L CurrentA5, (A0) ; this copies CurrentA5 into our local
; storage for global use in the listener
MOVE.W #DDPMaxData, D0 ; max size of data in a packet
_NewPtr CLEAR
BNE.S setUpFailed ; if NIL then forget it
LEA our_Buff, A1 ; we need to save the pointer reference
MOVE.L A0, (A1) ; in a place the listener can find it
MOVE.L A0, D0 ; return value to caller
RTS
setUpFailed
CLR.L D0 ; tell caller we failed by returning nil
; (caller expecting valid ptr returned)
RTS
ENDP
|
We now resume our regular programming...
You should typically get an AEP response packet within a few milliseconds. If
there is no response for a period of time, typically about 10 seconds, you
should resend your AEP request to account for a lost request or lost packets.
To be really safe, you should resend your AEP request with different data to
take into account the response to the first packet coming back later. The
retry time could then be simply set to k*Round_Trip_Time , where the
value of k depends upon the request semantics, like total data size.
This technique of statically setting the retry time is not always adequate to
accommodate the varying delays encountered in a internet environment at
different times. You could dynamically adjust the retry time based on an
adaptive retransmission algorithm that continuously monitors round trip times
and adjusts its timeout parameter accordingly. To implement an adaptive
algorithm, you can record the round trip time for each transaction. One common
technique is to keep the average round trip time as a weighted average and use
new round trip times from transactions to change the average slowly. For
example, one averaging technique** uses a constant weighing
factor, q , where 0 <= q < 1 , to weigh the oldest
average against the latest round trip time:
Choosing a value for q close to 1 makes the weighted average immune to
changes that last a short time. Choosing a value for q close to 0
makes the weighted average respond to changes in the delay very quickly.
The total time (i.e., retry time * retry count) before a request is
concluded as failed could be anywhere from 10 seconds to a couple of minutes,
depending on the type of the client application and the relative distance
between the source and the destination.
- **Douglass_Corner, InterNetworking with TCP/IP. KARN, P. and C. PARTRIDGE [August 1987], "Improving Round-Trip Time Estimates in Reliable Transport Protocols", Proceedings of ACM SIGCOMM 1987.
Back to top
NBP Retry Counts
You cannot really use the AEP to estimate round trip times for NBP packets
because you need to use NBP to determine the internet address of the node from
which an echo is being sought. In this case, you have to use the type of
device that you are looking for as a heuristic for setting the retry count.
The LaserWriter, for example, may be busy and not respond to a LkUp packet. In
such a case, you might want to do a quick lookup to return a partial list to
the user like the Chooser. You could then do a longer lookup to get a more
complete list of mappings. You should use a "back off" algorithm to make the
subsequent lookups further apart to generate progressively less traffic. Name
lookups are expensive and produce a lot of network traffic, and name
confirmation is the recommended call to use when confirming mappings obtained
through early bindings. Because Name lookups are expensive, you should avoid
searching all the zones in the internet.
Back to top
Setting TRel Timer in SendRequest
AppleTalk Phase 2 drivers allow you to set the TRel timer in
SendRequest or NSendRequest calls with ATP XO (exactly once)
service so as not to be locked into the pre-AppleTalk Phase 2 time of 30
seconds. You should set this timer based on the round trip time. Generally,
if the round trip time is less than one second, the default TRel time setting
of 30 seconds is adequate. If the round trip time is more, you can increase
the TRel time proportionately.
Back to top
xppTimeout and xppRetry
The two ZIP calls, GetZoneList and GetLocalZones , made on the
.XPP driver contain the ATP retry interval (in seconds) and count, in the
xppTimeout and xppRetry parameters. Both these functions are
ATP request-response transactions between a node and a router on the network to
which the requesting node is attached. The round trip is relatively short for
this transaction, and you should have very small values of xppTimeout
and xppRetry, typically two and three, respectively.
Back to top
References
Inside AppleTalk
Inside Macintosh, Volumes II & V, The AppleTalk Manager
M.NW.Internets
M.NW.AppleTalk2Mac
Back to top
Downloadables
|
Acrobat version of this Note (60K)
|
Download
|
Back to top
|