com.gemstone.org.jgroups.protocols.pbcast.STABLE.old Maven / Gradle / Ivy
Show all versions of gemfire-jgroups Show documentation
// $Id: STABLE.java,v 1.27 2005/07/13 19:57:59 belaban Exp $
package org.jgroups.protocols.pbcast;
import org.jgroups.*;
import org.jgroups.stack.Protocol;
import org.jgroups.util.Promise;
import org.jgroups.util.TimeScheduler;
import org.jgroups.util.Util;
import org.jgroups.util.Streamable;
import java.io.*;
import java.util.Properties;
import java.util.Vector;
import java.util.Iterator;
import java.util.Map;
/**
* Computes the broadcast messages that are stable, i.e. have been received by all members. Sends
* STABLE events up the stack when this is the case. This allows NAKACK to garbage collect messages that
* have been seen by all members.
* Works as follows: periodically we mcast our highest seqnos (seen for each member) to the group.
* A stability vector, which maintains the highest seqno for each member and initially contains no data,
* is updated when such a message is received. The entry for a member P is computed set to
* min(entry[P], digest[P]). When messages from all members have been received, a stability
* message is mcast, which causes all members to send a STABLE event up the stack (triggering garbage collection
* in the NAKACK layer).
* The stable task now terminates after max_num_gossips if no messages or view changes have been sent or received
* in the meantime. It will resume when messages are received. This effectively suspends sending superfluous
* STABLE messages in the face of no activity.
* New: when max_bytes is exceeded (unless disabled by setting it to 0),
* a STABLE task will be started (unless it is already running).
* @author Bela Ban
*/
public class STABLE extends Protocol {
Address local_addr=null;
final Vector mbrs=new Vector();
final Digest digest=new Digest(10); // keeps track of the highest seqnos from all members
final Promise digest_promise=new Promise(); // for fetching digest (from NAKACK layer)
final Vector heard_from=new Vector(); // keeps track of who we already heard from (STABLE_GOSSIP msgs)
long digest_timeout=60000; // time to wait until digest is received (from NAKACK)
/** Sends a STABLE gossip every 20 seconds on average. 0 disables gossipping of STABLE messages */
long desired_avg_gossip=20000;
/** delay before we send STABILITY msg (give others a change to send first). This should be set to a very
* small number (> 0 !) if max_bytes is used */
long stability_delay=6000;
StabilitySendTask stability_task=null;
final Object stability_mutex=new Object(); // to synchronize on stability_task
StableTask stable_task=null; // bcasts periodic STABLE message (added to timer below)
final Object stable_task_mutex=new Object(); // to sync on stable_task
TimeScheduler timer=null; // to send periodic STABLE msgs (and STABILITY messages)
int max_gossip_runs=3; // max. number of times the StableTask runs before terminating
int num_gossip_runs=max_gossip_runs; // this number is decremented (max_gossip_runs doesn't change)
static final String name="STABLE";
/** Total amount of bytes from incoming messages (default = 0 = disabled). When exceeded, a STABLE
* message will be broadcast and num_bytes_received reset to 0 . If this is > 0, then ideally
* stability_delay should be set to a low number as well */
long max_bytes=0;
/** The total number of bytes received from unicast and multicast messages */
long num_bytes_received=0;
/** When true, don't take part in garbage collection protocol: neither send STABLE messages nor
* handle STABILITY messages */
boolean suspended=false;
/** Max time we should hold off on message garbage collection. This is a second line of defense in case
* we get a SUSPEND_STABLE, but forget to send a corresponding RESUME_STABLE (which should never happen !)
* The consequence of a missing RESUME_STABLE would be that the group doesn't garbage collect stable
* messages anymore, eventually, with a lot of traffic, every member would accumulate messages and run
* out of memory !
*/
// long max_suspend_time=600000;
ResumeTask resume_task=null;
final Object resume_task_mutex=new Object();
/** Number of gossip messages */
int num_gossips=0;
public String getName() {
return name;
}
public long getDesiredAverageGossip() {
return desired_avg_gossip;
}
public void setDesiredAverageGossip(long gossip_interval) {
desired_avg_gossip=gossip_interval;
}
public long getMaxBytes() {
return max_bytes;
}
public void setMaxBytes(long max_bytes) {
this.max_bytes=max_bytes;
}
public int getNumberOfGossipMessages() {return num_gossips;}
public void resetStats() {
super.resetStats();
num_gossips=0;
}
public Vector requiredDownServices() {
Vector retval=new Vector();
retval.addElement(new Integer(Event.GET_DIGEST_STABLE)); // NAKACK layer
return retval;
}
public boolean setProperties(Properties props) {
String str;
super.setProperties(props);
str=props.getProperty("digest_timeout");
if(str != null) {
digest_timeout=Long.parseLong(str);
props.remove("digest_timeout");
}
str=props.getProperty("desired_avg_gossip");
if(str != null) {
desired_avg_gossip=Long.parseLong(str);
props.remove("desired_avg_gossip");
}
str=props.getProperty("stability_delay");
if(str != null) {
stability_delay=Long.parseLong(str);
props.remove("stability_delay");
}
str=props.getProperty("max_gossip_runs");
if(str != null) {
max_gossip_runs=Integer.parseInt(str);
num_gossip_runs=max_gossip_runs;
props.remove("max_gossip_runs");
}
str=props.getProperty("max_bytes");
if(str != null) {
max_bytes=Long.parseLong(str);
props.remove("max_bytes");
}
str=props.getProperty("max_suspend_time");
if(str != null) {
log.error("max_suspend_time is not supported any longer; please remove it (ignoring it)");
props.remove("max_suspend_time");
}
if(props.size() > 0) {
log.error("STABLE.setProperties(): these properties are not recognized: " + props);
return false;
}
return true;
}
void suspend(long timeout) {
if(!suspended) {
suspended=true;
if(log.isDebugEnabled())
log.debug("suspending message garbage collection");
}
startResumeTask(timeout); // will not start task if already running
}
void resume() {
suspended=false;
if(log.isDebugEnabled())
log.debug("resuming message garbage collection");
stopResumeTask();
}
public void start() throws Exception {
if(stack != null && stack.timer != null)
timer=stack.timer;
else
throw new Exception("STABLE.up(): timer cannot be retrieved from protocol stack");
}
public void stop() {
stopStableTask();
}
public void up(Event evt) {
Message msg;
StableHeader hdr;
Header obj;
int type=evt.getType();
switch(type) {
case Event.MSG:
msg=(Message)evt.getArg();
if(max_bytes > 0) { // message counting is enabled
long size=Math.max(msg.getLength(), 24);
num_bytes_received+=size;
if(num_bytes_received >= max_bytes) {
if(log.isTraceEnabled()) {
StringBuffer sb=new StringBuffer("max_bytes has been exceeded (max_bytes=");
sb.append(max_bytes).append(", number of bytes received=");
sb.append(num_bytes_received).append("): sending STABLE message");
log.trace(sb.toString());
}
new Thread() {
public void run() {
initialize();
sendStableMessage();
}
}.start();
num_bytes_received=0;
}
}
obj=msg.getHeader(name);
if(obj == null || !(obj instanceof StableHeader))
break;
hdr=(StableHeader)msg.removeHeader(name);
switch(hdr.type) {
case StableHeader.STABLE_GOSSIP:
handleStableGossip(msg.getSrc(), hdr.stableDigest);
break;
case StableHeader.STABILITY:
handleStabilityMessage(hdr.stableDigest);
break;
default:
if(log.isErrorEnabled()) log.error("StableHeader type " + hdr.type + " not known");
}
return; // don't pass STABLE or STABILITY messages up the stack
case Event.SET_LOCAL_ADDRESS:
local_addr=(Address)evt.getArg();
break;
}
passUp(evt);
if(desired_avg_gossip > 0) {
if(type == Event.VIEW_CHANGE || type == Event.MSG)
startStableTask(); // only starts task if not yet running
}
}
/**
* We need to receive this event out-of-band, otherwise we would block. The use case is
*
* - To send a STABLE_GOSSIP message we need the digest (from NAKACK below)
*
- We send a GET_DIGEST_STABLE event down from the up() method
*
- NAKACK sends the GET_DIGEST_STABLE_OK backup. However, we may have other messages in the
* up queue ahead of this event ! Therefore the event cannot be processed until all messages ahead of
* the event have been processed. These can't be processed, however, because the up() call waits for
* GET_DIGEST_STABLE_OK ! The up() call would always run into the timeout.
* @param evt
*/
protected void receiveUpEvent(Event evt) {
if(evt.getType() == Event.GET_DIGEST_STABLE_OK) {
digest_promise.setResult(evt.getArg());
return;
}
super.receiveUpEvent(evt);
}
public void down(Event evt) {
int type=evt.getType();
switch(evt.getType()) {
case Event.VIEW_CHANGE:
View v=(View)evt.getArg();
Vector tmp=v.getMembers();
mbrs.removeAllElements();
mbrs.addAll(tmp);
heard_from.retainAll(tmp); // removes all elements from heard_from that are not in new view
stopStableTask();
break;
case Event.SUSPEND_STABLE:
long timeout=0;
Object t=evt.getArg();
if(t != null && t instanceof Long)
timeout=((Long)t).longValue();
stopStableTask();
suspend(timeout);
break;
case Event.RESUME_STABLE:
resume();
break;
}
if(desired_avg_gossip > 0) {
if(type == Event.VIEW_CHANGE || type == Event.MSG)
startStableTask(); // only starts task if not yet running
}
passDown(evt);
}
public void runMessageGarbageCollection() {
sendStableMessage();
}
/* --------------------------------------- Private Methods ---------------------------------------- */
void initialize() {
synchronized(digest) {
digest.clear();
for(int i=0; i < mbrs.size(); i++) {
digest.add((Address)mbrs.elementAt(i), -1, -1);
// digest.add((Address)mbrs.elementAt(i), 0, 0);
}
heard_from.removeAllElements();
heard_from.addAll(mbrs);
}
}
void startStableTask() {
num_gossip_runs=max_gossip_runs;
// Here, double-checked locking works: we don't want to synchronize if the task already runs (which is the case
// 99% of the time). If stable_task gets nulled after the condition check, we return anyways, but just miss
// 1 cycle: on the next message or view, we will start the task
if(stable_task != null)
return;
synchronized(stable_task_mutex) {
if(stable_task != null && stable_task.running()) {
return; // already running
}
stable_task=new StableTask();
timer.add(stable_task, true); // fixed-rate scheduling
}
if(log.isTraceEnabled())
log.trace("stable task started; num_gossip_runs=" + num_gossip_runs + ", max_gossip_runs=" + max_gossip_runs);
}
void stopStableTask() {
// contrary to startStableTask(), we don't need double-checked locking here because this method is not
// called frequently
synchronized(stable_task_mutex) {
if(stable_task != null) {
stable_task.stop();
stable_task=null;
}
}
}
void startResumeTask(long max_suspend_time) {
max_suspend_time=(long)(max_suspend_time * 1.1); // little slack
synchronized(resume_task_mutex) {
if(resume_task != null && resume_task.running()) {
return; // already running
}
else {
resume_task=new ResumeTask(max_suspend_time);
timer.add(resume_task, true); // fixed-rate scheduling
}
}
if(log.isDebugEnabled())
log.debug("resume task started, max_suspend_time=" + max_suspend_time);
}
void stopResumeTask() {
synchronized(resume_task_mutex) {
if(resume_task != null) {
resume_task.stop();
resume_task=null;
}
}
}
void startStabilityTask(Digest d, long delay) {
synchronized(stability_mutex) {
if(stability_task != null && stability_task.running()) {
return; // already running
}
else {
stability_task=new StabilitySendTask(d, delay);
timer.add(stability_task, true); // fixed-rate scheduling
}
}
}
void stopStabilityTask() {
synchronized(stability_mutex) {
if(stability_task != null) {
stability_task.stop();
stability_task=null;
}
}
}
/**
Digest d contains (a) the highest seqnos deliverable for each sender and (b) the highest seqnos
seen for each member. (Difference: with 1,2,4,5, the highest seqno seen is 5, whereas the highest
seqno deliverable is 2). The minimum of all highest seqnos deliverable will be taken to send a stability
message, which results in garbage collection of messages lower than the ones in the stability vector. The
maximum of all seqnos will be taken to trigger possible retransmission of last missing seqno (see DESIGN
for details).
*/
private void handleStableGossip(Address sender, Digest d) {
Address mbr;
long highest_seqno, my_highest_seqno;
long highest_seen_seqno, my_highest_seen_seqno;
boolean my_own_gossip=false;
if(d == null || sender == null) {
if(log.isErrorEnabled()) log.error("digest or sender is null");
return;
}
if(suspended) {
if(log.isTraceEnabled()) {
log.trace("STABLE message will not be handled as suspended=" + suspended);
}
return;
}
if(log.isTraceEnabled())
log.trace(new StringBuffer("received digest from ").append(sender).append(": ").append(d));
if(!heard_from.contains(sender)) { // already received gossip from sender; discard it
if(log.isTraceEnabled()) log.trace("already received gossip from " + sender);
return;
}
// we won't handle the gossip d, if d's members don't match the membership in my own digest,
// this is part of the fix for the NAKACK problem (bugs #943480 and #938584)
if(!this.digest.sameSenders(d)) {
if(log.isTraceEnabled()) {
log.trace("received digest from " + sender + " (digest=" + d + ") which does not match my own digest ("+
this.digest + "): ignoring digest and re-initializing own digest");
}
initialize();
return;
}
my_own_gossip=local_addr != null && local_addr.equals(sender);
Map.Entry entry;
org.jgroups.protocols.pbcast.Digest.Entry val;
for(Iterator it=d.senders.entrySet().iterator(); it.hasNext();) {
entry=(Map.Entry)it.next();
mbr=(Address)entry.getKey();
if(!digest.contains(mbr)) {
if(log.isTraceEnabled()) log.trace("sender " + mbr + " not found in stability vector");
continue;
}
val=(org.jgroups.protocols.pbcast.Digest.Entry)entry.getValue();
highest_seqno=val.high_seqno;
highest_seen_seqno=val.high_seqno_seen;
// compute the minimum of the highest seqnos deliverable (for garbage collection)
my_highest_seqno=digest.highSeqnoAt(mbr);
if(my_highest_seqno < 0) {
if(highest_seqno >= 0 && my_own_gossip)
digest.setHighSeqnoAt(mbr, highest_seqno);
}
else {
digest.setHighSeqnoAt(mbr, Math.min(my_highest_seqno, highest_seqno));
}
// compute the maximum of the highest seqnos seen (for retransmission of last missing message)
my_highest_seen_seqno=digest.highSeqnoSeenAt(mbr);
if(my_highest_seen_seqno < 0) {
if(highest_seen_seqno >= 0)
digest.setHighSeqnoSeenAt(mbr, highest_seen_seqno);
}
else {
digest.setHighSeqnoSeenAt(mbr, Math.max(my_highest_seen_seqno, highest_seen_seqno));
}
}
heard_from.removeElement(sender);
if(heard_from.size() == 0) {
if(log.isTraceEnabled()) log.trace("sending stability msg " + digest);
sendStabilityMessage(digest.copy());
initialize();
}
}
/**
* Bcasts a STABLE message to all group members. Message contains highest seqnos of all members
* seen by this member. Highest seqnos are retrieved from the NAKACK layer above.
*/
void sendStableMessage() {
Digest d=null;
Message msg=new Message(); // mcast message
StableHeader hdr;
if(suspended) {
if(log.isTraceEnabled())
log.trace("will not send STABLE message as suspended=" + suspended);
return;
}
d=getDigest();
if(d != null && d.size() > 0) {
if(log.isTraceEnabled())
log.trace("mcasting STABLE msg, digest=" + d +
" (num_gossip_runs=" + num_gossip_runs + ", max_gossip_runs=" + max_gossip_runs + ')');
hdr=new StableHeader(StableHeader.STABLE_GOSSIP, d);
msg.putHeader(name, hdr);
num_gossips++;
passDown(new Event(Event.MSG, msg));
}
}
Digest getDigest() {
Digest ret=null;
passDown(new Event(Event.GET_DIGEST_STABLE));
ret=(Digest)digest_promise.getResult(digest_timeout);
if(ret == null) {
if(log.isErrorEnabled())
log.error("digest could not be fetched from below " + "(timeout was " + digest_timeout + " msecs)");
}
return ret;
}
/**
Schedules a stability message to be mcast after a random number of milliseconds (range 1-5 secs).
The reason for waiting a random amount of time is that, in the worst case, all members receive a
STABLE_GOSSIP message from the last outstanding member at the same time and would therefore mcast the
STABILITY message at the same time too. To avoid this, each member waits random N msecs. If, before N
elapses, some other member sent the STABILITY message, we just cancel our own message. If, during
waiting for N msecs to send STABILITY message S1, another STABILITY message S2 is to be sent, we just
discard S2.
@param tmp A copy of te stability digest, so we don't need to copy it again
*/
void sendStabilityMessage(Digest tmp) {
long delay;
if(timer == null) {
if(log.isErrorEnabled())
log.error("timer is null, cannot schedule stability message to be sent");
timer=stack != null ? stack.timer : null;
return;
}
// give other members a chance to mcast STABILITY message. if we receive STABILITY by the end of
// our random sleep, we will not send the STABILITY msg. this prevents that all mbrs mcast a
// STABILITY msg at the same time
delay=Util.random(stability_delay);
startStabilityTask(tmp, delay);
}
void handleStabilityMessage(Digest d) {
if(d == null) {
if(log.isErrorEnabled()) log.error("stability vector is null");
return;
}
if(suspended) {
if(log.isDebugEnabled()) {
log.debug("STABILITY message will not be handled as suspended=" + suspended);
}
return;
}
if(log.isDebugEnabled()) log.debug("stability vector is " + d.printHighSeqnos());
stopStabilityTask();
// we won't handle the gossip d, if d's members don't match the membership in my own digest,
// this is part of the fix for the NAKACK problem (bugs #943480 and #938584)
if(!this.digest.sameSenders(d)) {
if(log.isDebugEnabled()) {
log.debug("received digest (digest=" + d + ") which does not match my own digest ("+
this.digest + "): ignoring digest and re-initializing own digest");
}
initialize();
return;
}
// pass STABLE event down the stack, so NAKACK can garbage collect old messages
passDown(new Event(Event.STABLE, d));
}
/* ------------------------------------End of Private Methods ------------------------------------- */
public static class StableHeader extends Header implements Streamable {
public static final int STABLE_GOSSIP=1;
public static final int STABILITY=2;
int type=0;
// Digest digest=new Digest(); // used for both STABLE_GOSSIP and STABILITY message
Digest stableDigest=null; // changed by Bela April 4 2004
public StableHeader() {
} // used for externalizable
public StableHeader(int type, Digest digest) {
this.type=type;
this.stableDigest=digest;
}
static String type2String(int t) {
switch(t) {
case STABLE_GOSSIP:
return "STABLE_GOSSIP";
case STABILITY:
return "STABILITY";
default:
return "";
}
}
public String toString() {
StringBuffer sb=new StringBuffer();
sb.append('[');
sb.append(type2String(type));
sb.append("]: digest is ");
sb.append(stableDigest);
return sb.toString();
}
public void writeExternal(ObjectOutput out) throws IOException {
out.writeInt(type);
if(stableDigest == null) {
out.writeBoolean(false);
return;
}
out.writeBoolean(true);
stableDigest.writeExternal(out);
}
public void readExternal(ObjectInput in) throws IOException, ClassNotFoundException {
type=in.readInt();
boolean digest_not_null=in.readBoolean();
if(digest_not_null) {
stableDigest=new Digest();
stableDigest.readExternal(in);
}
}
public long size() {
long retval=Global.INT_SIZE + Global.BYTE_SIZE; // type + presence for digest
if(stableDigest != null)
retval+=stableDigest.serializedSize();
return retval;
}
public void writeTo(DataOutputStream out) throws IOException {
out.writeInt(type);
Util.writeStreamable(stableDigest, out);
}
public void readFrom(DataInputStream in) throws IOException, IllegalAccessException, InstantiationException {
type=in.readInt();
stableDigest=(Digest)Util.readStreamable(Digest.class, in);
}
}
/**
Mcast periodic STABLE message. Interval between sends varies. Terminates after num_gossip_runs is 0.
However, UP or DOWN messages will reset num_gossip_runs to max_gossip_runs. This has the effect that the
stable_send task terminates only after a period of time within which no messages were either sent
or received
*/
private class StableTask implements TimeScheduler.Task {
boolean stopped=false;
public void stop() {
stopped=true;
}
public boolean running() { // syntactic sugar
return !stopped;
}
public boolean cancelled() {
return stopped;
}
public long nextInterval() {
long interval=computeSleepTime();
if(interval <= 0)
return 10000;
else
return interval;
}
public void run() {
if(suspended) {
if(log.isTraceEnabled())
log.trace("stable task will not run as suspended=" + suspended);
stopStableTask();
return;
}
initialize();
sendStableMessage();
num_gossip_runs--;
if(num_gossip_runs <= 0) {
if(log.isTraceEnabled())
log.trace("stable task terminating (num_gossip_runs=" +
num_gossip_runs + ", max_gossip_runs=" + max_gossip_runs + ')');
stopStableTask();
}
}
long computeSleepTime() {
return getRandom((mbrs.size() * desired_avg_gossip * 2));
}
long getRandom(long range) {
return (long)((Math.random() * range) % range);
}
}
/**
* Multicasts a STABILITY message.
*/
private class StabilitySendTask implements TimeScheduler.Task {
Digest d=null;
boolean stopped=false;
long delay=2000;
public StabilitySendTask(Digest d, long delay) {
this.d=d;
this.delay=delay;
}
public boolean running() {
return !stopped;
}
public void stop() {
stopped=true;
}
public boolean cancelled() {
return stopped;
}
/** wait a random number of msecs (to give other a chance to send the STABILITY msg first) */
public long nextInterval() {
return delay;
}
public void run() {
Message msg;
StableHeader hdr;
if(suspended) {
if(log.isDebugEnabled()) {
log.debug("STABILITY message will not be sent as suspended=" + suspended);
}
stopped=true;
return;
}
if(d != null && !stopped) {
msg=new Message();
hdr=new StableHeader(StableHeader.STABILITY, d);
msg.putHeader(STABLE.name, hdr);
passDown(new Event(Event.MSG, msg));
d=null;
}
stopped=true; // run only once
}
}
private class ResumeTask implements TimeScheduler.Task {
boolean running=true;
long max_suspend_time=0;
ResumeTask(long max_suspend_time) {
this.max_suspend_time=max_suspend_time;
}
void stop() {
running=false;
}
public boolean running() {
return running;
}
public boolean cancelled() {
return running == false;
}
public long nextInterval() {
return max_suspend_time;
}
public void run() {
if(suspended)
log.warn("ResumeTask resumed message garbage collection - this should be done by a RESUME_STABLE event; " +
"check why this event was not received (or increase max_suspend_time for large state transfers)");
resume();
}
}
}