|
Corvettefisher posted:thumb rule 1 vCPU per gig of ram, I do 1 vcpu:2GB ram. You can assing 12 vCPU's but when contention time comes because 5 people decided it would be funny to download COUPON TOOL BAR that runs all CPU's at 100% you are really going to be wishing you limited them. Not only that but the more RAM/CPU assigned to a vm will create more overhead. A 12 core CPU will not perform anyfaster for the end user if 8 of those cores are idle, the scheduler will issue a wait for those 12 cores though causing some performance loss across the host. This isn't a rule of thumb at all... Typically you allocate 1 vCPU unless you know for a fact your application needs and will use more than 1 vCPU. There is no magic formula of X CPUs per gig of RAM. That said; the reason we don't assign 12 vCPUs to all VMs is that the rest WILL queue when one VM gets really busy. If you don't necessarily need to give the app 12 vCPUs to perform (maybe you only need 4?) then you will be better off assigning 4 vCPUs. This now means that for every 12 cores you can have 3 VMs running concurrently and will most likely get better overall performance out of the application. Here is a fantastic read: http://www.vmware.com/resources/techresources/10131 It's for 4.1 but not a lot has changed with the VMware scheduler between 4.X and 5.X. You'll know if things are going bad for you with assigning multiple vCPUs if you see %RDY climb and %CSTP climb in esxtop. The short of it is the less vCPUs a VM has/needs the higher consolidation ratio you'll be able to achieve. That 20%, can I get more context? Is it 20% of the total CPU resources in the system or is it 20% of one CPU core?
|
#
?
Aug 3, 2012 22:41
|
|
|
|
| # ? May 9, 2024 15:24 |
|
|
Corvettefisher posted:Depends how far you are over subscribing your hardware, how much those vcpu's are processing and how long they will be busy. They are desktop machines so lets imagine 1-2 cores 45sec out of every min and full on as-many-cores-as-I-can-get the other 15sec for an 8 hour workday. e: the workflow I'm evisioning is for rendering CG where the apps have become extremely parallel so the faster the artists can work interactively the better. The catch is that the heavy work comes very intermittantly. Move some sliders, check render, move sliders, etc. I'm guessing the downtime when adjustments are made account for 80% of the total time and 20% is spent waiting for results. That could wildly vary depending on how many cpus the vm has. Perhaps this workflow isn't very well suited for VMs but I'd like to dream. Shaocaholica fucked around with this message at 22:49 on Aug 3, 2012 |
|
#
?
Aug 3, 2012 22:43
|
|
|
Corvettefisher posted:thumb rule 1 vCPU per gig of ram, I do 1 vcpu:2GB ram. Every time I think you've made your dumbest post on this thread, you surprise me. The right way to do this is: 1 vCPU. Always. If you hit 100% on 1 vCPU regularly, bump it up to 2 vCPU. If you hit 100% on 2 vCPU regularly, bump it up to 3 vCPU. Repeat as necessary. If I gave every 4GB RAM VM of mine 2 vCPUs, I would probably double my vCPU count.
|
|
#
?
Aug 3, 2012 22:43
|
|
|
1000101 posted:This isn't a rule of thumb at all... No it isn't on official thumb rule, but it is one I generally use in VDI deployments, it seems to work pretty well. But in general only assign what you need. madsushi posted:Every time I think you've made your dumbest post on this thread, you surprise me. vCPU 1 and then bump it up is great in a static server environment. However for VDI's I found doing a 1:1 works best. Dilbert As FUCK fucked around with this message at 22:59 on Aug 3, 2012 |
|
#
?
Aug 3, 2012 22:45
|
|
|
Shaocaholica posted:They are desktop machines so lets imagine 1-2 cores 45sec out of every min and full on as-many-cores-as-I-can-get the other 15sec for an 8 hour workday. VMware is pretty bad at this type of thing. Honestly, one of its biggest weaknesses is handling bursty CPU load effectively. Until we can change core count on the fly, there's not an easy solution for this outside of just doing really low VM density on your physical hosts so that you can safely assign a lot of cores without worrying about contention.
|
|
#
?
Aug 3, 2012 22:51
|
|
|
Corvettefisher posted:No it isn't on official thumb rule, but it is one I generally use in VDI deployments, it seems to work pretty well. But in general only assign what you need. Stop using this rule, you're screwing up your consolidation ratio on a workload that traditionally rarely needs more than 1 vCPU anyway. You should be categorizing your users into a few buckets and then defining the VM specifications that support that bucket of users. For example you might have helpdesk people who have an email client open and a web browser and you might have sysadmins who use visio and a few java tools. Helpdesk peopleshould get 1 vCPU and maybe 2-3GB of RAM and the sysadmins MIGHT get 2 vCPUs and 4GB of RAM. Most likely they'll be fine with 1 vCPU and 4GB of RAM. The more vCPUs you give a VM the less the number of VMs you can concurrently schedule. Unless your users are pegging out the 1 vCPU you've assigned you're just wasting resources/money.
|
|
#
?
Aug 3, 2012 22:58
|
|
|
1000101 posted:Stop using this rule, you're screwing up your consolidation ratio on a workload that traditionally rarely needs more than 1 vCPU anyway. You should be categorizing your users into a few buckets and then defining the VM specifications that support that bucket of users. Yeah I tried 1vCPU even on the low end users, it was a complete headache at login, flash based website, or program start up. Saying "Don't try to open everything at once" doesn't cut it, because you get a backlash "Well I could do it on my physical computer". Doing 1:1 I have yet to see my CPU rise above 70% utilization on my hosts, and CPU ready is at a minimum. The max I ever do to a VDI is 4 cores, and as of yet complaints related to "lag" has dropped significantly. This deploy was with windows 7 areo on, normally I would have it windows classic but people often want a 'pretty' desktop
|
|
#
?
Aug 3, 2012 23:06
|
|
|
Zero clients are pretty cool! Some questions though. Are there any that support more than 1 display? Higher than WUXGA resolution per display? What about portable clients (laptops) that are not just regular laptops with a dongle so you can maximize battery life and weight?
|
|
#
?
Aug 3, 2012 23:08
|
|
|
Shaocaholica posted:Zero clients are pretty cool! Some questions though. Are there any that support more than 1 display? Higher than WUXGA resolution per display? What about portable clients (laptops) that are not just regular laptops with a dongle so you can maximize battery life and weight? My Samsung NC240 seems too, 1920x1200 as well. Andriod has a good app for vmware view however it is PCoIP only, so I would suggest an Asus Transformer Prime. Apple also has an app I believe.
|
|
#
?
Aug 3, 2012 23:10
|
|
|
Shaocaholica posted:Zero clients are pretty cool! Some questions though. Are there any that support more than 1 display? Higher than WUXGA resolution per display? What about portable clients (laptops) that are not just regular laptops with a dongle so you can maximize battery life and weight? From what I've seen in my little bit of looking, Teradici has some that will do 2x 1920x1200, soon to be 4x, and 10zig has something that will do 4x 1920x1200, but there's some weird thing you have to do, like two linked sessions or something. I think I'll be asking for some demo units soon, and we definitely need a minimum of 4x 1920x1200 here to make VDI viable.
|
|
#
?
Aug 3, 2012 23:41
|
|
|
Erwin posted:we definitely need a minimum of 4x 1920x1200 here to make VDI viable. You're crazier than me and I'm just playing all this up in my head. I don't actually deploy these things!
|
|
#
?
Aug 3, 2012 23:56
|
|
|
I get around 60 Windows 7 VMs (usually 1x3) on a 12 core HT DL380 with memory to spare from the 144Gb we provision them with as standard. We have an exceptionally resident-process heavy corporate desktop, so it's lower than usual, but no-one is doing complex graphical work. I still see CPU peg for the first 15 mins of any shift change.
|
|
#
?
Aug 4, 2012 01:59
|
|
|
Too bad most all my clients won't accept a vcpu at 100% for the first 15 minutes. 2vcpu's fixes this as only 33% log in at 7am, 33% log in at 11, and the others log in at 2. So far CPU contention has not been a problem, and 2 vcpu's has fixed most all complaints I get from VDI's slow. How many java based or in house programs do you run? GPO's or login scripts? How many programs do your users start up? Are you running windows Areo with transparency(full glass) or windows classic. Are most of the programs your clients used flash based? Are you using intel or AMD CPU's if so what clock speed?
|
|
#
?
Aug 4, 2012 02:34
|
|
|
Mausi posted:I get around 60 Windows 7 VMs (usually 1x3) on a 12 core HT DL380 with memory to spare from the 144Gb we provision them with as standard. What version of View and what's your storage look like?
|
|
#
?
Aug 4, 2012 02:37
|
|
|
If you use a view desktop with 1vCPU you are going to notice the all around sluggishness. At least on Windows 7. Lots of servers go into production at our shop with 1 vCPU but no desktops. The performance is not acceptable even for the security computers who only search the phonebook. This is with optimized by the book v5.0 linked clones.
|
|
#
?
Aug 4, 2012 04:09
|
|
|
Thank you for that post. VDI's are a completely different world than servers due to how dynamic they are.
|
|
#
?
Aug 4, 2012 05:48
|
|
|
Corvettefisher posted:Questions The hardware is HP DL380s running Intel E7450 @ 2.4Ghz, storage is primarily NetApp 6240s over 10GbE on Nexus at about 200% overallocation per volume. We never hit an IOPS issue on the filer, it's really just the desktop build is nowhere near optimised for virtualisation. And we don't use a connection broker, it's all 1:1 mappings, again for compliance. The 1x3s are for secondary Windows 7 VMs, we're exploring using 2x4 or 2x6 for primary machines for offshore workers, but it's pretty drat cost-inefficient because of the other platform limitations above.
|
|
#
?
Aug 4, 2012 10:15
|
|
|
Erwin posted:From what I've seen in my little bit of looking, Teradici has some that will do 2x 1920x1200, soon to be 4x, and 10zig has something that will do 4x 1920x1200, but there's some weird thing you have to do, like two linked sessions or something. I think I'll be asking for some demo units soon, and we definitely need a minimum of 4x 1920x1200 here to make VDI viable. I have tested the Teradici I9440 and it worked pretty well but wouldn't recommend it over any sort of WAN link. We ultimately went with a Wyse model that has 2x single-link DVI ports, so 1920x12 max. They work fine. Wyse's packaging and attention to detail sucks: Each zero-client came in a PC desktop sized box, and it came with a PS/2 mouse (no PS/2 ports on the zero client).
|
|
#
?
Aug 4, 2012 22:25
|
|
|
Is antivirus on Virtualcenter a known "don't do that" ? I'm pretty sure I just discovered our recent VMWare cluster outages (can't http to any ESXi nodes, VMs just drop off entirely) are caused by Microsoft Forefront on the Virtualcenter server. feld fucked around with this message at 04:29 on Aug 7, 2012 |
|
#
?
Aug 7, 2012 04:14
|
|
|
I noticed today that some of our Linux guests have zero swap space provisioned within the OS. That's...not good, right? Or does the hypervisor completely take over swapping duties? Running under vSphere 4.0 (yeah I know) if it matters.
|
|
#
?
Aug 9, 2012 21:24
|
|
|
Docjowles posted:I noticed today that some of our Linux guests have zero swap space provisioned within the OS. That's...not good, right? Or does the hypervisor completely take over swapping duties? Running under vSphere 4.0 (yeah I know) if it matters. Hypervisor swapping is dumb swapping. It'll throw down to disk whatever it can grab, whether it's accessed a lot or not. The OS will be much smarter about what to swap, so you should definitely have it do so. The hypervisor won't swap unless it needs to, and first it will try to use the balloon driver to force the OS to swap more intelligently. So also make sure that VMware tools is installed.
|
|
#
?
Aug 9, 2012 21:29
|
|
|
There's a few things you need to really, really keep in mind here. First, the hypervisor's first line of defense against memory overcommitment is ballooning. When ESXi starts to run out of memory, the balloon driver in VMware Tools will coerce the operating system to give back somewhere in the neighborhood of 35% (configurable) of its available memory to the hypervisor. In order to do that, it needs to page out to disk. If it can't page out to disk, your hypervisor will run out of memory and start doing much worse things like doing memory compression (in 4.1+) or hypervisor swap. Second, hypervisor swap is far more impactful on performance than host-side swap. In fact, most host-side swap is pretty harmless -- the operating system tends to swap things out when they sit idle in memory for too long anyway. But when a process/thread needs to swap in from disk, the thread hangs idle while it waits for the operating system to finish dealing with the page fault. When the hypervisor needs to swap in, it means hanging the entire virtual machine instance. Expect network latency to be through the roof and throughput on everything, not just memory-bound tasks, to be frustratingly slow.
|
|
#
?
Aug 9, 2012 22:28
|
|
|
Thanks for the info. The good news is, if I'm reading the perf charts in vCenter right, it's not swapping at all. Swap used, swap in and swap out are all 0. But it doesn't give me the warm fuzzies that at any moment performance could go to poo poo due to something as dumb as not setting up a swap file. We do appear to at least be using VMware tools, which I wasn't taking for granted at first.
|
|
#
?
Aug 9, 2012 22:47
|
|
|
One of my customers has decided that he wants to roll out thin clients. It's a medical facility where thin clients would help security and minimize impact of hardware being damaged so I can't argue with the idea, but I don't have a clue where to start. Some form of VDI, Citrix, or MSTS seem to be my main options. Can you guys point me to any good resources for determining which is the best fit, how much it'll cost, and how to properly size the servers? Right now I don't see any reason to go beyond MSTS, most of the users do nothing but interact with web-based apps (Chromebooks were at one point seriously discussed), but I know enough to know where I need to know more, and this is one of those cases. I know TS/Citrix isn't entirely within the scope of this thread, but I'm pretty sure I've seen it discussed here a few times and this seems to be the closest one available.
|
|
#
?
Aug 10, 2012 15:58
|
|
|
Find a consultant that's done this before. I've looked at VDI and Thin Clients before and there's just so many gotchas in setting up an environment you really need a skilled consultant that's done it before. Last time I really looked at the numbers, Citrix was the most expensive. VMWare View is pretty nice these days, or you can go straight Microsoft Remote Desktop Services (RDS is the new name for Terminal Services). MS VDI with all the new R2 improvements is supposedly not terrible and much cheaper than laying VMWare or Citrix licenses over the MS licenses. Sizing the back end is the hardest part. If 30 nurses all show up and log on at 7AM you need to make sure the backend doesn't poo poo itself.
|
|
#
?
Aug 10, 2012 16:46
|
|
|
I wish I could just go to a consultant. This is one of those boss-driven projects, where the customer in question is a friend of his and the boss wants to get in to managing their data network rather than just the phones as we do now. He wants to offer the data management at a fairly low price, so hiring an outsider is entirely out of the question. I have been and will continue to make the argument that this is something that's over our heads and should be left to others, but he sees another customer (who he's also trying to get in to managing data for) who has a Citrix system being managed by possibly the most retarded IT company I've ever dealt with where things haven't exploded horribly, so he assumes if they can do it then we can as well. Citrix vs. MSTS/RDS/whatever is obviously the easiest one, since Citrix lays on top of MSTS and requires the same licensing as a base, the question is does Citrix add anything to the equation that's worth their additional licensing cost. So far what I've come across does not lead me to believe this is the case given all the RDS improvements in 2008. Their largest advantage claimed for simple uses like this site is having easy clients for many devices. Last time I checked I had an RDP client on everything I own that resembles a computer, so that doesn't seem like a selling point to me. VDI vs. TS is the more open question, I think VDI would require more resources but would make the setup and management more like deploying ordinary desktops which we're all familiar with rather than everyone actually running on the same server. I just don't know if the additional server cost and licensing for the client OSes and possibly VM platform would be worth the more "traditional" environment. Of course it's all irrelevant if we can't figure out sizing. tl;dr: ugh. wolrah fucked around with this message at 18:31 on Aug 10, 2012 |
|
#
?
Aug 10, 2012 18:29
|
|
|
how many people, and what kind of workload are you looking at?
|
|
#
?
Aug 10, 2012 18:36
|
|
|
Anytime you are dealing with virtualizing the user's workspace, it depends on the apps. That is the single most important question. You need to take a good look at the apps and how they use them. With the limited amount of information that we have, it sounds to me like TS/RDS is your best bet. Are they going to be connecting from outside of the network? Working remotely, etc?
|
|
#
?
Aug 10, 2012 18:52
|
|
|
Two sites, around 40-50 users at any given time, 24/7 operation, and the vast majority of users are using nothing more than a web browser to access a variety of "cloud" services. About a third use an office suite, where there's a mix of MS Office and OpenOffice depending on how much money the guy had when he was buying computers. A half dozen or so use more than that and those few may be keeping their normal computers. So unless the web services have some terrible Java/Flash interface or go way overboard on the Javascript it should be a reasonably low load. The servers I have currently in my preliminary spec are a pair of Dell T410s with dual Xeon E5620s (2.4GHz quad core) and 24GB of RAM a piece which met the recommendations of some sizing guide I found and since lost. It seems to me like there should be a fair bit of headroom on such a system, but I'll be the first to admit I'm basically guessing here. I do think I under-specced my disks, not for space but I/O capacity, it's only set for 4x 1TB 7200 RPM drives. edit: Remote access may happen, but likely only for the fairly tech-savvy owner. wolrah fucked around with this message at 18:56 on Aug 10, 2012 |
|
#
?
Aug 10, 2012 18:54
|
|
|
How are you guys going to handle the mix of Office Suites and licensing? Also keep in mind if those are OEM licenses of Windows XP you cannot just virtualize them for VDI. From what you have said I'd probably go about it by making those two servers virtualization hosts and then put TS/RDS servers on top. I use Citrix and have never worked with the base TS/RDS, so I can't offer a whole lot of advice there. I would enable Desktop Experience and just publish a desktop for each user. Are you planning on using those disks as file storage? If those are only virtualization hosts you should be fine as long as you are giving the servers enough memory. [Edit: Any sort of sizing guide will be fairly useless as it all depends on the apps.]
|
|
#
?
Aug 10, 2012 19:03
|
|
|
The owner is fine with going full OpenOffice, he uses it himself and says the MS Office installs are only still there because they're paid for and work. They use Google Docs for most internal office suite needs and only keep the local clients for the occasional more advanced feature. We will not be virtualizing XP, most of them are Home edition currently anyways, any VDI will be with new licenses if we go that way. The disks in the systems are for the system, existing file servers will be remaining for users.
|
|
#
?
Aug 10, 2012 19:30
|
|
|
Anyone have anything to share on the performance impact of changing storage data counter levels in 5.0 U1? I'm experiencing the "no data available" bug on historical datastore performance graphs after updating to 5.0 U1, and apparently some hacky PowerCLI nonsense is required to get them working again.
|
|
#
?
Aug 11, 2012 22:44
|
|
|
Can anyone guess why my host-attached storage is so slow in esxi5? I'm running on an old IBM x3650 (m0) with two 146GB sas-drives in raid1 (i think). 24GB ram, only one Xeon E5345@2.33GHz-cpu. The storage adapter is an ServeRAID 8k/8k-l8. Esxi runs from usb-stick. This server is NOT on the hcl, not for esxi5 anyway. It also does not support passthrough. I tried running this at my HOST # date && dd if=/dev/zero bs=1024 count=1000000 of=dd.file && date Mon Aug 13 07:28:02 UTC 2012 1000000+0 records in 1000000+0 records out Mon Aug 13 07:30:18 UTC 2012 (I'm no *nix guru) but it should mean that writing 1GB of data started at 07:28:02 and finished at 07:30:18 which means it took two minutes 16 seconds? Running the same at home, in a GUEST bsd where the host is an old laptop with 5400 or 7200rpm sata-drive running server 2003 x86 and virtualization provided by vmware workstation version old, Mån 13 Aug 2012 09:29:06 CEST 1000000+0 records in 1000000+0 records out 1024000000 bytes transferred in 32.284491 secs (31718016 bytes/sec) Mån 13 Aug 2012 09:29:39 CEST The same command on the same laptophost but in another guest in virtualbox instead takes 62 seconds. Even if it's not on the hcl, why would a server-class machine, with sas-drives, run so insanely much slower than what a laptop does? There is 0 room to purchase "SAN", the only "NAS" we have is a consumer-grade zyxel pos where copying an iso to/from the esxi-server took hours. Nothing makes any sense to me, the only thing i know is that something isn't right. (This is my last week at this place so i probably won't care to "fix" this, it runs fine for the crap it's supposed to do, i'm just curious if i've unchecked something that should be checked, or if i've done something that is the biggest NO-NO or what..)
|
|
#
?
Aug 13, 2012 08:49
|
|
|
underlig posted:Can anyone guess why my host-attached storage is so slow in esxi5? What firmware is the ServeRAID adapter running? I've encountered multiple issues with ServeRAID 8k adapters in the past which were resolved by upgrading the firmware. Also there is a known issue with the SAS-expander backplane on the x3650s although that is only supposed to affect SATA HDDs. Just make a bootable update disk with IBM BoMC and use that to update the firmware of all components. Edit: I've just noticed that you might need to use v3.00 of BoMC as they seem to have discontinued support for all M0 System x servers in the newer versions.
|
|
#
?
Aug 13, 2012 13:29
|
|
|
dd is not a good indicator of performance due to imposed limits for console resources. You'll want to test performance using a VM with appropriate benchmarking software. If you really want to use the console, you can do some tests like creating eager zeroed thick VMDK files and time that. E.g.: `time vmkfstools -c 1G test.vmdk -d eagerzeroedthick` Testing reads might still be an issue for the console. Really, just use a VM! Edit: Wrong reply button... plus removed some derpiness. Kachunkachunk fucked around with this message at 13:40 on Aug 13, 2012 |
|
#
?
Aug 13, 2012 13:36
|
|
|
cheese-cube posted:What firmware is the ServeRAID adapter running? I've encountered multiple issues with ServeRAID 8k adapters in the past which were resolved by upgrading the firmware. Also there is a known issue with the SAS-expander backplane on the x3650s although that is only supposed to affect SATA HDDs. Will try to update anyway and see what happens. What could possibly go wrong..  Kachunkachunk posted:dd is not a good indicator of performance due to imposed limits for console resources. You'll want to test performance using a VM with appropriate benchmarking software. But these imposed limits, they're still there if i say copy a vmdk from vsphere client right? I mean it's the same slow deal there? I remember when i had problems resizing a vmdk and did a copy of it before testing, it took hours.
|
|
#
?
Aug 13, 2012 13:53
|
|
|
Copying files in your datastore browser should use the full speeds possible, but keep in mind that's a read+write task, so expect it to be somewhat slower than pure [edit: "reads"] of course. I'm pretty sure you meant otherwise though. So probably it's best you try to use esxtop on the box to measure latency and go from there. There's a good guide on Yellow Bricks that explains a lot of the metrics available. You should look at your DAVG values: http://www.yellow-bricks.com/esxtop/ Edited. Somehow I missed a word? Weird. Kachunkachunk fucked around with this message at 07:29 on Aug 15, 2012 |
|
#
?
Aug 14, 2012 03:30
|
|
|
Veeam want to come in and sell me stuff - I've already got Netuitive for Performance/Capacity/Rightsizing and my backup/replication isn't going anywhere near their platform, so what the hell else do they have? Maybe reporting or chargeback modelling - but I'm pretty sure I can do Chargeback with Netuitive if I wanted to, and I've already got an in-house app for environment reporting.
|
|
#
?
Aug 14, 2012 21:55
|
|
|
So have them wine and dine you instead of coming in? don't do this
|
|
#
?
Aug 14, 2012 22:31
|
|
|
|
| # ? May 9, 2024 15:24 |
|
|
1000101 posted:I thought I would post this for the 3 of you out there actually running your vCenter database on Oracle and are about to upgrade from 4.1 to 5.0 update 1: vCenter 5.0 Update 1b was released yesterday, including a resolution to this issue. Yeah, we run vCenter on Oracle too.
|
|
#
?
Aug 17, 2012 15:07
|
|
