Thursday, September 01, 2011

Optimizations for large clusters

I've been working on making JGroups more efficient on large clusters. 'Large' is between 100 and 2000 nodes.

My focus has been on making the memory footprint smaller, and to reduce the wire size of certain types of messages.

Here are some of the optimizations that I implemented.


Discovery is needed by a new member to find the coordinator when joining. It broadcasts a discovery request, and everybody in the cluster replies with a discovery response.

There were 2 problems with this: first, a cluster of 1000 nodes meant that a new joiner received 1000 messages at the same time, possibly clogging up network queues and causing messages to get dropped.

This was solved by staggering the sending of responses (stagger_timeout).

The second problem was that every discovery response included the current view. In a cluster of 1000, this meant that 1000 responses each contained a view of 1000 members !

The solution to this was that we only send back the address of the coordinator; as this is all that's needed to send a JOIN request to it. So instead of sending back (with every discovery response) 1000 addresses, we now only send back 1 address.


A digest used to contain the lowest, highest delivered and highest received sequence numbers (seqnos) for every member. They are sent back to a new joiner in a JOIN response, and they are also broadcast periodically by STABLE to purge messages delivered by everyone.

The wire size would be 2 longs for every address (UUID), and 3 longs for the 3 seqnos. That's roughly 1000 * 5 * 8 = 40000 bytes for a cluster of 1000 members. Bear in mind that that's the size of one digest; in a cluster of 1000, everyone broadcasts such a digest periodically (STABLE) !

The first optimization was to remove the 'low' seqno; I had to change some code in the retransmitters to allow for that, but - hey - who wouldn't do that to save 8 bytes / STABLE message ? :-)

This reduced the wire (and memory !) size of a 1000-member digest by another 8'000 bytes, down to 32'000 (from 40'000).

Having only highest delivered (HD) and highest received (HR) seqnos allowed for another optimization: HR is always >= HD, and the difference between HR and HD is usually small.

So the next optimization was to send HR as a delta to HD. So instead of sending 322649 | 322650, we'd send 322649 | 1.

The central optimization underlying that was that seqnos seldomly need 8 bytes: a seqno starts at 1 and increases monotonically. If a member sends 5 million messages, the seqno can still be encoded in 4 bytes (saving 4 bytes per seqno). If a member is restarted, the seqno starts again at 1 and can thus be encoded in 1 byte.

So now I could encode an HD/HR pair by sending a byte containing the number of bytes needed for the HD part in the lower 4 bits and the number of bytes needed for the delta in the higher 4 bits. The HD and the delta would then follow. Example: to encode HD=2000000 | HR=2000500, we'd generate the bytes:

| 50 | -128 | -124 | 30 | -12 | 1 |

  • 50 encodes a length of 3 for HD and 2 for HD-HR (500)
  • -128, -124 and 30 encode 2'000'000 in 3 bytes
  • -12 and 1 encode the delta (500)

So instead of using 16 bytes for the above sequence, we use only 6 bytes !

If we assume that we can encode 2 seqnos on average in 6 bytes, the wire size of a digest is now 1000 * (16 (UUID) + 6) = 22'000, that's down from 40'000 in a 1000 member cluster. In other words, we're saving almost 50% of the wire size of a digest !

Of course, we can not only encode seqno sequences, but also other longs, which is exactly what we did for another optimization. Examples of where this makes sense are:
  • Seqnos in NakackHeaders: every multicast message has such a header, so the savings here are significant
  • Range: this is used for retransmission requests, and is also a seqno sequence
  • RequestCorrelator IDs: used for every RPC
  • Fragmentation IDs (FRAG and FRAG2)
  • UNICAST and UNICAST2: sqnos and ranges
  • ViewId
An example of where this doesn't make sense are UUIDs: they are generated such that the bits are spread out over the entire 8 bytes, so encoding them would make 9 bytes out of 8 and that doesn't help.


A JoinRsp used to contain a list of members twice: once in the view and once in the digest. The was eliminated, and now we're sending the member list only once. This also cut the wire size of a JoinRsp in half.

Further optimizations planned for 3.1 include delta views and better compressed STABLE messages:

Delta views

If we have a view of 1000 members, we always send the full address list with every view change. This is not necessary, as everybody has access to the previous view.

So, for example, when we have P, Q and R joining, and X and Y leaving in V22, then we can simply send a delta view; a view V22={V21+P+Q+R-X-Y}. This means, take the current view V21, remove members X and Y, and add members P, Q and R to the tail of the list, in order to generate a new view V22.

So, instead of sending a list of 1000 members, we simply send 5 members, and everybody creates the new view locally, based on the current view and the delta information.

Compressed STABLE messages

A STABLE message contains a digest with a list of all members and then the digest seqnos for HD and HR. Since STABLE messages are exchanged between members of the same cluster, they all have the same view, or else they would drop a STABLE message.

Hence, we can drop the View and instead send the ViewId, which is 1 address and a long. Everyone knows that the digest seqnos will be in order of the current view, e.g. seqno pair 1 belongs to the first member of the current view, seqno pair 2 to the second member and so on.

So instead of sending a list of 1000 members for a STABLE message, we only send 1 address.

This will reduce the wire size of a 1000-member digest sent by STABLE from roughly 40'000 bytes to ca. 6'000 bytes !

Download 3.0.0.CR1

The optimizations (exluding delta views and compressed STABLE messages) are available in JGroups 3.0.0.CR1, which can be downloaded from [1].

Enjoy (and feedback appreciated, on the mailing lists...) !



  1. Hi,
    I have run a cluster with 104 nodes on 13 servers with 8 core (using JGroups for network communication) and the application works perfectly without any problem.
    Now we are going to extends the number of servers,so we will have 736 nodes. I didn't change The JGroups configuration for the new cluster. I tried to start the cluster but only 202 nodes joined to the cluster. I send the configuration for u and please let know what is wrong with. The JGroups version is 2.10.0.GA.
    Thank you very much

  2. 736 servers, wow !

    As I mentioned in a previous blog post, I doubt you can run more than 500-1000 nodes with a standard configuration, but need to make onfiguration changes, or upgrade to a newer JGroups version.
    I suggest (a) upgrade to 2.12.1.Final (2.12.2 coming soon)and (b) use udp-largecluster.xml. This configuration has changes compared to udp.xml which are geared toward running in large clusters.

  3. vahid1:08 PM

    Thanks for your quick answer,
    Well, I will try it. Maybe it is better to mention We are working on 46 servers.Each one has 16 cores and with 16 JGroups instances running on it.
    736(nodes) = 46(servers) * 16(core)

    Anyway thank you,

  4. Vahid1:26 PM

    I have defined a master node that sends requests to workers and get the results from the worker nodes. I want to reduce the network traffic by optimizing the JGroups. I want nodes to communicate only with the master node(no communication with each other). Is there any way to restrict nodes communication ?


  5. No, there isn't; JGroups is a peer-to-peer system, and the nodes need to talk to each other

  6. Hi Bela,

    Can you suggest how to write application using JGroups to run it in multiple boxes so that node can interact with each other as in cluster.