Coda File System

Next Previous Contents

15. Failure Detection

Three hazards face the user of an RPC package: @begin(Enumerate) The communication medium may fail.

The peer process at a remote site may crash. @end(Enumerate)

A key problem in RPC is reliably detecting either of these events when an RPC call is in progress. Detection of failures in the absence of RPC calls in progress is an orthogonal issue, and can be reduced to this issue by generating artificial keepalive RPC calls.

Ideally, the detection of these failures should be independent of the specific RPC call in progress. In other words, as long as we are sure that communication medium is not broken and that the remote server process is alive, we should not care how long it takes to receive the reply to an RPC request. At the same time failures should be detected as soon as possible, so that suitable recovery actions can be performed. The following paragraphs show this goal is achieved in RPC2.

15.1 Retransmission Algorithm

When the RPC2 runtime system receives a retry packet for a request it is already working on, it responds with a Busy packet. There are two constants B@-[total] and N . These constants are set in @RPC2(Init)()], with suitable defaults built in. These semantics of these two constants are: @begin(Enumerate) Communication failure is declared if N successive retries of a packet fail to provoke any kind of response. The response may be a reply, a Busy packet, an acknowledgement if the packet being sent is a reply, or an implicit piggy-backed acknowledgement.

Site failure is declared if silence is observed for a total period of time in the range B @-[ total ] to 2B @-[total ]. @end(Enumerate)

RPC2 does not try to accurately distinguish between site failure and communication failure: one may masquerade as the other, and a single failure @RPC2(DEAD) reflects both cases. Loosely speaking, N characterises the probability of packet loss in the communication medium, while B @-[ total ] characterises how sluggish a server may get before it is declared dead.

Given B@-[total] and N , we can determine B@-[1], B@-[2], ... B@-[N] such that @w{ B@-[1] + B@-[2] + B@-[3] ... B@-[N] } = @w{ B@-[total] and B@-[i] < B@-[i+1 .] } Each B@-[i] is a retry interval and the progressive lengthening of these intervals is to allow for transient overloads at remote sites. In RPC2, B @-[ i+i ] = 2B @-[ i ]. In practise we may place a minimum bound on the values for B @-[ i ]s, to avoid send out packets too close to each other.

The RPC2 packet transmission algorithm is based on these concepts and is outlined as follows: @begin(ProgramExample)

while (TRUE) { for (i = 0; i < N; i++) { send (packet); awaitresponse (B@-[i]); if (reply or lastack arrived) quit; if (BUSY arrived) break; } if (i > = N) goto TimeOut; sleep (B@-[total]); }

TimeOut: mark connection @RPC2(DEAD); mark dead all other connections to this (host, portal) pair; @end(ProgramExample)

Failure is detected in time B]@-[total] if the remote site dies just after the sleep () call ends. If the failure occurs immediately after the remote site sends a Busy packet, failure is detected after a total of 2B@-[total .] These cases bound the time it takes to detect failure. Failure is also declared if all N of the retries are lost due to communication failure. This will occur in a time exactly equal to B@-[total] .

How does this mesh with side effects? The above algorithm will work regardless of the duration of a side effect as long as there is some response from the server in RPC (a Busy packet ) or the side effect (any packet) at intervals of B @-[ total ]. Note that it is immaterial whether the side effect involves asynchronous Unix processes or not. If such processes are involved their failure will be detected (perhaps as @RPC2(DEAD) failures or in other ways) and reported by the remote server explicitly as @RPC2(SEFAIL2). Only if the remote server is itself dead or unreachable is the RPC return code @RPC2(DEAD) and this will occur no later than 2B @-[ total ] after the failure. In SFTP or where TCP or other protocols are being used for side effects, the failure detection mechanisms of these protocols will be relied upon to detect side effect failure.

The tables at the end of this chapter show how the N retransmissions take place within B@-[total] , for typical values of N and B@-[total] . The original attempt is at time 0. The numbers in parentheses indicate the time (@i{B@-[N]}) that RPC2 waits after the transmission of the last retry, before declaring failure. A lower limit of 300 milliseconds for the retry interval is assumed.

Adjusting Retransmission Intervals

Retransmission intervals in RPC are maintained on a connection-specific basis. They are initialized with the values in the tables, given the number of retries and total timeout specified in RPC2_Init . There is a minimum bound of 300 milliseconds for the B@-[i] so that packets are not sent out too close to each other. For slower networks, this may still be too short an interval. The minimum interval should be at least a packet round trip time (RTT). To estimate the minimum interval, an RPC client collects RTT observations on bind1-bind2, bind3-bind4, initmulticast-reply, and request-reply pairs. A "smoothed RTT" is calculated (a la TCP) and used to set the lower limit on RPC retransmission intervals. Whenever the RTT estimate is changed, the retransmission intervals are recalculated, always maintaining the given timeout. The maximum value for the lower limit is @g(b)0, so the client has time for at least one retry.

RTT observations are collected by timestamping packets as described for TCP in @cite(RFC1323). The client stamps an outgoing packet with the current time. The server echoes that timestamp on the next packet back to that client. The client computes the RTT observation by subtracting the echoed stamp from the current time.

Two longwords in the packet header are used to carry timestamps. @footnote(RPC uses one for collecting RTT observations. It uses the other field to send the time for the bind sequence to the server. The bind time is used to initialize side effects. SFTP uses both timestamp fields. This is described further in chapter XXX .) Time units must be meaningful for the range 10Mbps - 1200bps. To keep the stamp small, time is measured relative to connection creation. The resolution of the stamp is 10 msec. Though it may be simpler to maintain a finer granularity unit (the maximum packet size is 2900 bytes, which over Ethernet takes 2.32 ms to transmit), the clock resolution on most machines is 10-20 ms. The 10 msec unit is still finer than TCP, which uses 500 ms ticks. A connection can stay open over a year before the time "wraps around" using 10 msec units.

The RTT and RTT variance are stored scaled, as in TCP, for greater precision and ease of computation. We need a longword for each, unlike TCP (which uses a short). If we used a short, and scaled by 8, we would have only 12 bits for time (40.950 sec). This is roughly the RTT of a packet over a 1200 bps link: it takes 19.33 sec to transmit a 2900 byte packet.

For compatibility, if a packet arrives with a 0 timestamp, it means the sender of the packet does not maintain or use RTT estimates. RPC simply skips the RTT update, and the retransmission timeout stays at its original setting.

Timestamps should be set as late as possible and collected as early as possible. Packets leaving the client are stamped in rpc2_SendReliably (for multicast, in mrpc_SendPacketReliably ). Retries are restamped. The RTT is updated and retransmission intervals adjusted from the socket listener, upon receipt of a bind2, bind4, or response.

Packets arriving at the server are checked in the SocketListener . If the request is a "good" init1, init3, or request, the timestamp is stashed in the connection entry and the request time is taken. The reply is stamped with the original timestamp plus the service time.

Requests that provoke busies are stamped with the incoming timestamp -- the service time is assumed to be 0. The client gets additional observations this way. Erroneous requests and binds that will be rejected are sent back with null timestamps. In such cases, there is usually no connection state to update.

RPC does not currently take request and reply length into account in its RTT estimate. Because of this, the RTT has high variance at low bandwidths. Ideally, RPC would derive an RTT independent of length, then apply it given the length of the request and probable length of the reply.

@newpage @string(Included="Yes") @include(retry.mss)

Next Previous Contents