TheServerSide Java Symposium 2007 - Day 3 Report


News: TheServerSide Java Symposium 2007 - Day 3 Report

  1. Day 1 | Day 2 A Chat With Azul Azul Systems participated in the Server Side Java Symposium as both exhibitor and speakers. I had a chance to sit down for ten minutes to speak with Gil Tene, CTO and cofounder at Azul. "I've been blown away with what we have accomplished in our generational pauseless garbage collection performance," Gil said, "Using Kirk Pepperdine's performance benchmark we have measure 35 Gbytes/second of continuous, no gc pause, hashmap allocation rates!" Gil told me Cliff Click has a JavaOne talk on a log3 hashtable that they plan to open source. It scales linearly to 700 cpus or more. Gil said he expect 10 Gbytes/second and was "blown away by 35." By comparison the typical Linux system clocks at 1 Gbytes/second. This is pauseless garbage collection and a key strength of the Azul system. Azul is now shipping multi-cpu servers – 96 being a small number of CPUs – and Gil said they are clearly focused on their sales efforts, even if it means some of their tremendous technical innovations wait a little longer to be brought to market. For instance, Gil said "It's not good enough to build a machine that runs a lot of little things, we need to let big things run in our environment." They see a key advantage for Azul's customers in providing dynamic ranging. Azul is learning to speak the customer's language now that they are in the market. In one experience a customer told him that they solved the garbage collection pause problem by tuning the gc. Gil said it is possible to fine tune the gc but there is no way around the several-times-a-day minutes-long gc pause without pauseless garbage collection. In another instance a customer running multiple instances of Tangasol Coherence found that a 5, 10, or 15 second gc pause causes Coherence to treat the node as crashed. Eventually Coherence will bring up the node again but this can happen all across a Coherence deployment. Gil said he now listens for customers who use 64-bit VMs. He said "Using a 64 bit VM is a symptom. You're using 64-bit because you need a bigger heap." Gil told me that one of Azul's challenges is to enable customers to know that there is a solution to gc pauses. Gil said sometimes customers "only find problems where they know solutions exist." Along the same lines Azul is now fighting a reputation for delivering a proprietary solution. In reality the Azul system runs Java applications from boxes running Linux, Solaris, Mac OS X, and applications like SAP and Peoplesoft. His goal is to show what developers experience on Azul, where you should be using Azul, and explaining the Azul systems and how they are used. Gil told me they are considering creating a special version of SpecEJB. Gil notes that SpecEJB has no response time requirements. Azul is toying with a SpecEJB that has no transaction times greater.5 second. "Let's see someone else run at that speed!" Gil said. Oracle Buys Tangosol With the creases of a new Oracle shirt that had just been pulled out of a box still visible, Cameron Purdy of Tangosol strode forth at the Symposium as an employee of Oracle. On a press conference call Oracle described the purchase as a way to add 5000 sales people to the Tangosol effort. Tangosol will be part of Oracle's Hot Pluggable strategy to allow customers applications to plug-in without changes. Oracle intends to make the Coherence product available as a stand-alone product and also as an option for Oracle Fusion Middleware. No financial terms were disclosed. (Which made me wonder why RedHat disclosed all the gory details of its purchase of JBoss.) Oracle expects to close the purchase in April 2007. Details are found at Oracle said they expect pricing to remain the same and while the Oracle license is different from the existing Tangosol license they are almost identical and will be working through the differences. Oracle expects to announce more details on licensing after the acquisition closes. The Tangosol team will go under the name Oracle Coherence and Oracle intends to retain all employees and current commitments. That translates to "the Coherence party at JavaOne is still on!" Oracle acknowledged that some Tangosol partners compete with Oracle but they had nothing to say about how they would resolve the conflicts. XML and Performance I gave the XML Performance talk this afternoon. The slides are now available on the Symposium Wiki. Frank Cohen
  2. Coherence timeouts[ Go to top ]

    In another instance a customer running multiple instances of Tangasol Coherence found that a 5, 10, or 15 second gc pause causes Coherence to treat the node as crashed.
    The node is automatically failed over (no lost data or transactions) once a configured timeout is hit. At that point, Coherence doesn't care if it crashed or not, and when it's done GC-ing, the load fails back. By default, I think the configurable timeout is 30 seconds. However, the real problem is that transactions that might usually take 5ms are now taking 10 seconds (for the GC) plus 5ms, which is 2000x longer due to the GC!!! So Azul's pauseless GC is definitely a great tool for eliminating latency spikes for this type of processing. Peace, Cameron Purdy Tangosol Coherence: The Java Data Grid