[ovirt-users] Re: Tuning and testing GlusterFS performance

Hey, thanks! Good catch! Going to have to take a look at that, will be working on it this weekend.. hopefully we can do this post creation. Thanks again Bill *From:* Jayme <jaymef(a)gmail.com&gt; *Sent:* Thursday, August 2, 2018 5:56 PM *To:* William Dossett <william.dossett(a)gmail.com&gt; *Cc:* users <users(a)ovirt.org&gt; *Subject:* Re: [ovirt-users] Tuning and testing GlusterFS performance Bill, I thought I'd let you (and others know this) as it might save you some headaches. I found that my performance problem was resolved by clicking "optimize for virt store" option in the volume settings of the hosted engine (for the data volume). Doing this one change has increased my I/O performance by 10x alone. I don't know why this would not be set or recommended by default but I'm glad I found it! - James On Thu, Aug 2, 2018 at 2:32 PM, William Dossett <william.dossett(a)gmail.com&gt; wrote: Yeah, I am just ramping up here, but this project is mostly on my own time and money, hence no SSDs for Gluster… I’ve already blown close to $500 of my own money on 10Gb ethernet cards and SFPs on ebay as my company frowns on us getting good deals for equipment on ebay and would rather go to their preferred supplier – where $500 wouldn’t even buy half a 10Gb CNA ☹ but I believe in this project and it feels like it is getting ready for showtime – if I can demo this in a few weeks and get some interest I’ll be asking them to reimburse me, that’s for sure! Hopefully going to get some of the other work off my plate and work on this later this afternoon, will let you know any findings. Regards Bill *From:* Jayme <jaymef(a)gmail.com&gt; *Sent:* Thursday, August 2, 2018 11:07 AM *To:* William Dossett <william.dossett(a)gmail.com&gt; *Cc:* users <users(a)ovirt.org&gt; *Subject:* Re: [ovirt-users] Tuning and testing GlusterFS performance Bill, Appreciate the feedback and would be interested to hear some of your results. I'm a bit worried about what i'm seeing so far on a very stock 3 node HCI setup. 8mb/sec on that dd test mentioned in the original post from within a VM (which may be explained by bad testing methods or some other configuration considerations).. but what is more worrisome to me is that I tried another dd test to time creating a 32GB file, it was taking a long time so I exited the process and the VM basically locked up on me, I couldn't access it or the console and eventually had to do a hard shutdown of the VM to recover. I don't plan to host many VMs, probably around 15. They aren't super demanding servers but some do read/write big directories such as working with github repos and large node_module folders, rsyncs of fairly large dirs etc. I'm definitely going to have to do a lot more testing before I can be assured enough to put any important VMs on this cluster. - James On Thu, Aug 2, 2018 at 1:54 PM, William Dossett <william.dossett(a)gmail.com&gt; wrote: I usually look at IOPs using IOMeter… you usually want several workers running reads and writes in different threads at the same time. You can run Dynamo on a Linux instance and then connect it to a window GUI running IOMeter to give you stats. I was getting around 250 IOPs on JBOD sata 7200rpm drives which isn’t bad for cheap and cheerful sata drives. As I said, I’ve worked with HCI in VMware now for a couple of years, intensely this last year when we had some defective Dell hardware and trying to diagnose the problem. Since then the hardware has been completely replaced with all flash solution. So when I got the all flash solution I used IOmeter on it and was only getting around 3000 IOPs on enterprise flash disks… not exactly stellar, but OK for one VM. The trick there was the scale out. There is a VMware Fling call HCI Bench. Its very cool in that you spin up one VM and then it spawns 40 more VMs across the cluster. I could then use VSAN observer and it showed my hosts were actually doing 30K IOPs on average which is absolutely stellar performance. Anyway, moral of the story there was that your one VM may seem like its quick, but not what you would expect from flash… but as you add more VMs in the cluster and they are all doing workloads, it scales out beautifully and the read/write speed does not slow down as you add more loads. I’m hoping that’s what we are going to see with Gluster. Also, you are using mb nomenclature below, is that Mb, or MB? I am sort of assuming MB megabytes per second… it does not seem very fast. I’m probably not going to get to work more on my cluster today as I’ve got other projects that I need to get done on time, but I want to try and get some templates up and running and do some more testing either tomorrow or this weekend and see what I get in just basic writing MB/s and let you know. Regards Bill *From:* Jayme <jaymef(a)gmail.com&gt; *Sent:* Thursday, August 2, 2018 8:12 AM *To:* users <users(a)ovirt.org&gt; *Subject:* [ovirt-users] Tuning and testing GlusterFS performance So I've finally completed my first HCI build using the below configuration: 3x Dell PowerEdge R720 2x 2.9 GHz 8 Core E5-2690 256GB RAM 2x250gb SSD Raid 1 (boot/os) 2x2TB SSD jbod passthrough (used for gluster bricks) 1Gbe Nic for management 10Gbe nic for Gluster Using Replica 3 with no arbiter. Installed the latest version of oVirt available at the time 4.2.5. Created recommended volumes (with an additional data volume on second SSD). Not using VDO First thing I did was setup glusterFS network on 10Gbe and set it to be used for glusterFS and migration traffic. I've setup a single test VM using Centos7 minimal on the default "x-large instance" profile. Within this VM if I do very basic write test using something like: dd bs=1M count=256 if=/dev/zero of=test conv=fdatasync I'm seeing quite slow speeds, only 8mb/sec. If I do the same from one of the hosts gluster mounts i.e. host1: /rhev/data-center/mnt/glusterSD/HOST:data I get about 30mb/sec (which still seems fairly low?) Am I testing incorrectly here? Is there anything I should be tuning on the Gluster volumes to increase performance with SSDs? Where can I find out where the bottle neck is here, or is this expected performance of Gluster?