Re: NETTYPE

Just a quick thought that the 8200 has only 2 physical cpu's so by the book that allows 1 cpu vp A far better approach would be to use the set up Art suggests with a 3rd cpu in the 8200 We have noticed over a number of years that the third cpu and therefore 2nd cpu vp makes the system run far smoother. Regards Nick Leonard Systems Administrator Poole Hospital NHS Trust >>> "Art S. Kagel" <kagel@bloomberg.net> 01/31/00 10:32pm >>> mars1972@my-deja.com wrote: > > In article <61CEA8293505D211A0E90000F87AAD7D2E02EE@srrcex1.srrpa.es>, > =?iso-8859-1?Q?Pablo_Herrero_Fern=E1ndez?= <PabloF.Herrero@srrpa.es> > wrote: > > Hello everybody. > > We have a DEC-UX Alpha8200 with 2 CPU and Informix 7.31. I have the > next > > values in ONCONFIG: > > > > NETTYPE soctcp,1,50,NET # Configure poll thread(s) for nettype > > NETTYPE ipcshm,2,200,CPU # Configure poll thread(s) for> nettype > > RESIDENT 1 # Forced residency flag (Yes = 1, No = > 0) > > > > MULTIPROCESSOR 1 # 0 for single-processor, 1 for> > multi-processor > > NUMCPUVPS 3 # Number of user (cpu) vps > > SINGLE_CPU_VP 0 # If non-zero, limit number of cpu> vps to > > one > > > > NOAGE 0 # Process aging > > AFF_SPROC 0 # Affinity start processor > > AFF_NPROCS 0 # Affinity number of processors> > > > # Shared Memory Parameters > > > > LOCKS 100000 # Maximum number of locks > > BUFFERS 53248 # Maximum number of shared buffers > > NUMAIOVPS 2 # Number of IO vps > > SHMBASE 0x200000000 # Shared memory base address > > SHMVIRTSIZE 155648 # initial virtual shared memory> segment size > > SHMADD 20480 # Size of new shared memory segments > > (Kbytes) > > SHMTOTAL 0 # Total shared memory (Kbytes). > 0=>unlimited> > > > Now, we'll have more users through soctcp (>400) than ipcshm (<30) > > connections. How can i set the values of NETTYPE? > > > This is what I would do to start... > > NETTYPE soctcp,4,200,CPU Problem: Pablo only has 3CPU VPs so the above will allocate at least one NET VP to handle the fourth listener. > NETTYPE ipcshm,1,50,NET> > This will allow up to 800 users over tcp, and 50 via shared memory, > using CPU class VPs for the tcp connections. > > However, (this is for you, Obnoxio), some people disagree with me on the > use of CPU VPs for tcp connections. What I gave you is the > recommendataion that I got from an Informix Performance Tuning class, > but you might want to switch them around and see if there's a difference > for you. I'm the big disagree-er on this issue (I guess on most I'm just contrary :-). If you put sharec memory listeners in NET VPs they will have to poll, spin the CPUs, and flood the OS with system calls to perform their task. If you put TCP listeners in the CPU VPs they also will have to poll the socket so they will also hog the CPU. However, if you place NET listeners in NET VPs they can block on a select() system call until there is a connect request or data block and then wake and process the new connection or request. If you place the shared memory connections in CPU VPs (and use all 3 CPU VPs) then since the CPU VPs cannot block, since they have work to do, they will periodically poll the shared memory blocks for connections and new data and will be able to handle the requests quickly. If all the CPU VPs are listening then requests do not have to wait for CPU VP#1, which is the busiest always, to become free to schedule the new request. So with my prior post the NETTYPES become: NETTYPE soctcp,2,250,NET # 2 x 250 = 500 connections supported. NETTYPE ipcshm,3,50,CPU # 50 connections is the default so might as well. Art S. Kagel