Re: Informix ODS 7.24 on Sequent NUMA-Q

In article <3502d2f5.0@news.snip.net>, David K. Killough <killougd@ix.netcom.com> wrote: >We are having problems testing a 450-user OLTP application. We have been >told that the Informix/NUMA-Q combination will not scale well for our test. David, Whoever told you that doesn't like Sequent and/or Informix :) Did this person told you why it wouldn't scale? >Anyone with INFORMIX/NUMA-Q OLTP installed please respond with comments!!! Please see my comments inserted below. >We have 2 quads (1GB mem, 4 180MHz P6 on each). We would like them to run >our 4GL based application with 450 users at 40% utilized (CPU). We have >seen 310 users at 70% (with less than acceptable thruput). Both quads were >being used. Quad 0 was running our engine, and a run queue was set up to >pull everything else to Quad 1. Is it possible that there is too much >inter-quad activity?? We have seen Symmetry equipment do much better >scaling with OLTP. You do not mention scaling here. What is your reference point? Did you actually run the very same application with the very same user simulation on both Symmetry on NUMA-Q? If so, I can tell you what you have seen: For OLTP workloads on the Informix Dynamic Server, a NUMA-Q quad is the rough equivalent of 8 to 10 Symmetry recent processors, depending on the application. I have personnally measured this ratio. The ratio can even be as good as 12, but that was not on an OLTP workload. Of course the ratio may vary depending on the application. In this particular case, may I suggest that you increase the number of Informix CPU VPs to 5 (4 on quad 1, 1 on quad 0). The ideal number of Informix CPU VPs is usually 5 for 2-quad systems and OLTP workload, although sometimes it's 4 indeed. (On 3 quads use 6-7 CPU VPs, on 4 quads use 8-9) Typically, the RDBMS itself uses more CPU than the 4GL processes. Allocate 60%-70% of your NUMA-Q CPUs to the RDBMS and 30%-40% to the 4GL users. >Our application will need to support 850 users by year 2000. We have 550 >now. Is NUMA-Q techology capable of OLTP at these levels?? Shouldn't be any problem with that. The current combination (IDS v7.24+ Pentium-Pro-based NUMA-Q) is handling thousands of OLTP users in production. Your Sequent and/or Informix representative will kindly provide you with actual references if you wish so. Of course, the number of users itself is meaningless as it depends on your application. You claim here that you max out 2 quads with 310 users, they must be pretty busy :) I've seen another OLTP application where we pack 450-500 users per quad (100-125 users per Pentium Pro). Do you have actual users typing or do you use a simulation software? If it's the latter, make sure you think time is realistic. You have to understand that you won't know how your application is scaling until you run it on several hardware configurations (1 quad, 2 quads, 3 quads, etc.). If you have test results on a Symmetry platform (for the same application and same Informix version) please contact me (see signature) and I'll help you extrapolating to NUMA-Q. >To better understand the capabilities of each quad without the complexities >of inter-quad accesses, we will probably shut down Quad 1 and test Quad 0 >all alone. We could partition some of the application but the majority >(70%) must remain together. Doing so would allow you to measure scaling from 1 quad to 2 quads. The scaling efficiency of the combination Informix Dynamic Server v7.24.UC1/ Sequent NUMA-Q (Pentium-Pro based) is 90% (from 1 quad to 2 quads) on an OLTP workload (based on actual customer benchmark). Of course, the Hardware/OS and the RDBMS are not the only factors that influence scalability: sometime scalability issues are caused by the application itself, usually a contention on a particular RDBMS table or on a resource shared by all users (lock, shared memory counter, file, etc.) If you max out two quads with 310 users, you should max out one quad with 170 users (roughly). If it's not what you see, a serious look at your application design, RDBMS configuration, and basic NUMA-Q tuning parameters are needed. >Also, Informix has claimed that ODS 7.24 (but not 7.23) is capable of >scaling. I have searched the release notes without much luck. IDS v7.24 contains a first set of improvements and v7.30 contains a second one. These improvements have a positive effect on both scalability and pure performance. For instance, I've seen IDS v7.24 service 30% more users than IDS v7.23 for the same SMP hardware (Symmetry). Please contact me directly (see signature) if you are interested by the measured scaling efficiencies of v7.22, v7.23, v7.24 and v7.30 on NUMA-Q platforms. I can have this sent to your Sequent or Informix representative in no time. To quote a benchmark result, the scaling efficiency of IDS v7.30 from 2 quads to 4 quads is 95%, even better than v7.24 (scaling efficiency of v7.30 on smaller configurations, such as from 1 quad to 2 quads, was not measured). This is for an OLTP workload, of course. Hope this helps, -- Laurent P. Montaron RDBMS Software engineering manager Email: lpm@sequent.com PGP key available, keyID 4E238995 Phone: +1 (503) 578-4395 Fax: + 1 (503) 578-3131 Sequent Computer Systems, 15450 SW Koll Pkwy, Beaverton OR 97006, USA -- Laurent P. Montaron RDBMS Software engineering manager Email: lpm@sequent.com PGP key available, keyID 4E238995 Phone: +1 (503) 578-4395 Fax: + 1 (503) 578-3131 Sequent Computer Systems, 15450 SW Koll Pkwy, Beaverton OR 97006, USA