Re: Channel Contention
Thanks, I wasn't aware of that. It makes the channel activity displays much less interesting. Brian Nielsen On Thu, 18 Feb 2010 20:12:57 -0500, Raymond Higgs rayhi...@us.ibm.com wrote: I wouldn't rely on those System Activity Display numbers with big IOs to determine channel capacity. That 3-4% is how busy the powerpc processor is in the channel. With big IOs, it's just idling, waiting on the DMA engines. If you add more IOs to the channel, you'll max out some piece of hardware way before the SAD panel shows 50%. With one of my 8 gig FCP channels, I see 25% at max read bandwidth. I think ficon would yield similar numbers. The processor in 2 and 4 gig channels is the same. The processor in 8 g ig channels is quite a bit faster. Ray Higgs System z FCP Development Bld. 706, B24 2455 South Road Poughkeepsie, NY 12601 (845) 435-8666, T/L 295-8666 rayhi...@us.ibm.com
Re: Channel Contention
Being that this is a one-time effort, I wouldn't worry too much about where the bottleneck is because you indicate can't do anything about it anyway. Better is simply to estimate the amount of time it will take to some gross level of precision. If you were going to be doing it regularl y then it becomes a more important question and could drive configuration changes. I know tape drives are not involved in your process, but I do know that DDR is very effective at driving the channel to them. We have a bank of 4 3590's shared between VM and z/OS on 2 ESCON channels. When I had 2 DDR' s writing to 2 of the tapes drives it effectivley saturated the channels from the VM LPAR and the z/OS jobs trying to use the other 2 tape drives suffered horribly. The reports from Velocity's ESAMAP made it easy to diagnose this cross-LPAR interference. (I ended up moving my DDR's to a different time slot.) I mention this to say that it will be good that there will not be other production workload going on to the DASD when you do this. BTW, the 3-4% channel utilization I mentioned in my other post is on 2G FICON channels. Obviously, 4G or 8G channels would make a big difference . Brian Nielsen On Wed, 17 Feb 2010 16:45:23 -0800, Schuh, Richard rsc...@visa.com wrot e: I will have to check with the TPF and H/W folks to get answers to most of your questions - VM was not included in the planning for the install. Since this is a one-time effort, I doubt that I can justify any new feature. I am fairly certain that the Ficon to the target disks is 4G. We are on a z10 (one I can answer authoritatively). The configuration of the switch is one that I absolutely cannot answer until I get an answer from the h/w folks. Unfortunately, they are in an earlier time zone, so I cannot get the answer until tomorrow.
Re: Channel Contention
The IBM z/VM Operating System IBMVM@LISTSERV.UARK.EDU wrote on 02/18/2010 10:48:52 AM: Brian Nielsen bniel...@sco.idaho.gov Sent by: The IBM z/VM Operating System IBMVM@LISTSERV.UARK.EDU 02/18/2010 10:48 AM Please respond to The IBM z/VM Operating System IBMVM@LISTSERV.UARK.EDU To IBMVM@LISTSERV.UARK.EDU cc Subject Re: Channel Contention Being that this is a one-time effort, I wouldn't worry too much about where the bottleneck is because you indicate can't do anything about it anyway. Better is simply to estimate the amount of time it will take to some gross level of precision. If you were going to be doing it regularl y then it becomes a more important question and could drive configuration changes. I know tape drives are not involved in your process, but I do know that DDR is very effective at driving the channel to them. We have a bank of 4 3590's shared between VM and z/OS on 2 ESCON channels. When I had 2 DDR' s writing to 2 of the tapes drives it effectivley saturated the channels from the VM LPAR and the z/OS jobs trying to use the other 2 tape drives suffered horribly. The reports from Velocity's ESAMAP made it easy to diagnose this cross-LPAR interference. (I ended up moving my DDR's to a different time slot.) I mention this to say that it will be good that there will not be other production workload going on to the DASD when you do this. BTW, the 3-4% channel utilization I mentioned in my other post is on 2G FICON channels. Obviously, 4G or 8G channels would make a big difference . Brian Nielsen I wouldn't rely on those System Activity Display numbers with big IOs to determine channel capacity. That 3-4% is how busy the powerpc processor is in the channel. With big IOs, it's just idling, waiting on the DMA engines. If you add more IOs to the channel, you'll max out some piece of hardware way before the SAD panel shows 50%. With one of my 8 gig FCP channels, I see 25% at max read bandwidth. I think ficon would yield similar numbers. The processor in 2 and 4 gig channels is the same. The processor in 8 gig channels is quite a bit faster. Ray Higgs System z FCP Development Bld. 706, B24 2455 South Road Poughkeepsie, NY 12601 (845) 435-8666, T/L 295-8666 rayhi...@us.ibm.com
Channel Contention
Currently, we have 3 LPARS, 2 support Linux and 1 for TPF testing. The current disk configuration is * a boatload of big 3390s (27-32MB) are on the Linux LPARs, These are connected using 4 Ficon channels. The DDRs will be done from one of the Linux LPARs. There will be two concurrent DDRs for this. These will be full disk copies. 4 channels to 210 disks. * Another boatload of 3330-03s where the TPF test system base disks reside. These are connected to the third LPAR via 8 ESCON channels to each array. Since the disks are not the same size, the minidisks will be copied, not the physical disks. The plan is to have 16 concurrent copies. There are 16 channels serving 437 disks, roughly 15,000 minidisks. * The target disks are connected via 4 Ficon channels that are EMIFd to all LPARs. There is separation of arrays; the TPF and Linux disks are not intermingled; however, the same 4 channels are shared between the LPARs. There are only 4 channels. The question is, will the 4 channels be a bottleneck if both the Linux and TPF migrations are done concurrently? Regards, Richard Schuh
Re: Channel Contention
Here are some data points that may help you. When I run standalone DDR to do a DASD-to-DASD copy within a single ESS 800 connected with 4 FICON channels, each channel is about 3-4% busy as reported on the HMC channel activity display, and the CPU is about 1% busy. My quick back of the envelope calculation is that about 25 concurrent DDR's would max out my 4 channels. If your source and targets are on different channels you could estimate getting twice that many before saturating 4 FICON channels. However, my other experiences with mass I/O tells me that the cache in your target DASD controller is also a good candidate to be the bottleneck. Initially the writes will be fast until the controller cache fills up, then it will slow down once it becomes neccessary to wait for tracks to be destaged from cache to disk. Therefore, while the amount of cache in the controller is a critical performance factor in normal workloads, in this situation however much controller cache you have will sooner or later get swamped. At that point you should check your DASD hardware performance guide for it's maximum sustained write throughput rates. Compare that to your channel throughput capacity to decide which one is the bottleneck. Depending on how much data you push, you may even find that the bottleneck starts at the channel and then shifts to the controller after the cache fills up. Brian Nielsen On Wed, 17 Feb 2010 11:14:27 -0800, Schuh, Richard rsc...@visa.com wrot e: Currently, we have 3 LPARS, 2 support Linux and 1 for TPF testing. The current disk configuration is * a boatload of big 3390s (27-32MB) are on the Linux LPARs, These are connected using 4 Ficon channels. The DDRs will be done from one of the Linux LPARs. There will be two concurrent DDRs for this. These will b e full disk copies. 4 channels to 210 disks. * Another boatload of 3330-03s where the TPF test system base disk s reside. These are connected to the third LPAR via 8 ESCON channels to eac h array. Since the disks are not the same size, the minidisks will be copied, not the physical disks. The plan is to have 16 concurrent copies. There are 16 channels serving 437 disks, roughly 15,000 minidisks. * The target disks are connected via 4 Ficon channels that are EMIFd to all LPARs. There is separation of arrays; the TPF and Linux disk s are not intermingled; however, the same 4 channels are shared between the LPARs. There are only 4 channels. The question is, will the 4 channels be a bottleneck if both the Linux and TPF migrations are done concurrently? Regards, Richard Schuh
Re: Channel Contention
I might be more concerned that the ESCON channels might be a bottle neck. Some of the factors that will affect the performance on the new DASD would be the speed of the Ficon channels (2 gig, 4 gig or 8 gig) and the internal performance of the new DASD compared to the old DASD. Paul Feller AIT Mainframe Technical Support From: The IBM z/VM Operating System [mailto:ib...@listserv.uark.edu] On Behalf Of Schuh, Richard Sent: Wednesday, February 17, 2010 1:14 PM To: IBMVM@LISTSERV.UARK.EDU Subject: Channel Contention Currently, we have 3 LPARS, 2 support Linux and 1 for TPF testing. The current disk configuration is * a boatload of big 3390s (27-32MB) are on the Linux LPARs, These are connected using 4 Ficon channels. The DDRs will be done from one of the Linux LPARs. There will be two concurrent DDRs for this. These will be full disk copies. 4 channels to 210 disks. * Another boatload of 3330-03s where the TPF test system base disks reside. These are connected to the third LPAR via 8 ESCON channels to each array. Since the disks are not the same size, the minidisks will be copied, not the physical disks. The plan is to have 16 concurrent copies. There are 16 channels serving 437 disks, roughly 15,000 minidisks. * The target disks are connected via 4 Ficon channels that are EMIFd to all LPARs. There is separation of arrays; the TPF and Linux disks are not intermingled; however, the same 4 channels are shared between the LPARs. There are only 4 channels. The question is, will the 4 channels be a bottleneck if both the Linux and TPF migrations are done concurrently? Regards, Richard Schuh
Re: Channel Contention
Either 4(most likely) or 8 gig, new dasd is faster and has more cache. I think they got 8g for the production TPF systems and 4 for VM. Regards, Richard Schuh From: The IBM z/VM Operating System [mailto:ib...@listserv.uark.edu] On Behalf Of Feller, Paul Sent: Wednesday, February 17, 2010 1:32 PM To: IBMVM@LISTSERV.UARK.EDU Subject: Re: Channel Contention I might be more concerned that the ESCON channels might be a bottle neck. Some of the factors that will affect the performance on the new DASD would be the speed of the Ficon channels (2 gig, 4 gig or 8 gig) and the internal performance of the new DASD compared to the old DASD. Paul Feller AIT Mainframe Technical Support From: The IBM z/VM Operating System [mailto:ib...@listserv.uark.edu] On Behalf Of Schuh, Richard Sent: Wednesday, February 17, 2010 1:14 PM To: IBMVM@LISTSERV.UARK.EDU Subject: Channel Contention Currently, we have 3 LPARS, 2 support Linux and 1 for TPF testing. The current disk configuration is * a boatload of big 3390s (27-32MB) are on the Linux LPARs, These are connected using 4 Ficon channels. The DDRs will be done from one of the Linux LPARs. There will be two concurrent DDRs for this. These will be full disk copies. 4 channels to 210 disks. * Another boatload of 3330-03s where the TPF test system base disks reside. These are connected to the third LPAR via 8 ESCON channels to each array. Since the disks are not the same size, the minidisks will be copied, not the physical disks. The plan is to have 16 concurrent copies. There are 16 channels serving 437 disks, roughly 15,000 minidisks. * The target disks are connected via 4 Ficon channels that are EMIFd to all LPARs. There is separation of arrays; the TPF and Linux disks are not intermingled; however, the same 4 channels are shared between the LPARs. There are only 4 channels. The question is, will the 4 channels be a bottleneck if both the Linux and TPF migrations are done concurrently? Regards, Richard Schuh
Re: Channel Contention
I will have to check with the TPF and H/W folks to get answers to most of your questions - VM was not included in the planning for the install. Since this is a one-time effort, I doubt that I can justify any new feature. I am fairly certain that the Ficon to the target disks is 4G. We are on a z10 (one I can answer authoritatively). The configuration of the switch is one that I absolutely cannot answer until I get an answer from the h/w folks. Unfortunately, they are in an earlier time zone, so I cannot get the answer until tomorrow. Regards, Richard Schuh From: The IBM z/VM Operating System [mailto:ib...@listserv.uark.edu] On Behalf Of Mike Rydberg Sent: Wednesday, February 17, 2010 11:58 AM To: IBMVM@LISTSERV.UARK.EDU Subject: Re: Channel Contention Richard, A couple question: How is your FICON Switch fabric configured, is this a single switch or cascaded switches between the channels and your DASD CU's? What is the link speed of your FICON Channels, 2Gb, 4Gb or 8Gb? If the LPAR's are running on a z10 you might benefit from the zHPF (high performance FICON feature on the z10), which reduces the number of bi-directional exchanges for each I/O. Also, depending on the FICON Switches in your shop, you should be able to monitor the FICON channel port utilization real time by using the web-based or CLI based switch port performance displays (eg. Brocade portperfshow cmd). I can give you a few tips if you wish to contact me offline. Regards Mike Rydberg Brocade From: The IBM z/VM Operating System [mailto:ib...@listserv.uark.edu] On Behalf Of Schuh, Richard Sent: Wednesday, February 17, 2010 1:14 PM To: IBMVM@LISTSERV.UARK.EDU Subject: Channel Contention Currently, we have 3 LPARS, 2 support Linux and 1 for TPF testing. The current disk configuration is * a boatload of big 3390s (27-32MB) are on the Linux LPARs, These are connected using 4 Ficon channels. The DDRs will be done from one of the Linux LPARs. There will be two concurrent DDRs for this. These will be full disk copies. 4 channels to 210 disks. * Another boatload of 3330-03s where the TPF test system base disks reside. These are connected to the third LPAR via 8 ESCON channels to each array. Since the disks are not the same size, the minidisks will be copied, not the physical disks. The plan is to have 16 concurrent copies. There are 16 channels serving 437 disks, roughly 15,000 minidisks. * The target disks are connected via 4 Ficon channels that are EMIFd to all LPARs. There is separation of arrays; the TPF and Linux disks are not intermingled; however, the same 4 channels are shared between the LPARs. There are only 4 channels. The question is, will the 4 channels be a bottleneck if both the Linux and TPF migrations are done concurrently? Regards, Richard Schuh