From: Joseph L. <jo...@ge...> - 2015-08-18 13:22:27
|
Solution: Put the linux client on the right network, and it works now. Rest of my email that I was writing up before I realized what I’d done: Sorry for the confusion, the performance issue was on a FreeBSD client with the FreeBSD servers, which became my drive to test some other clients (linux, mac). Both the linux and mac clients do the same thing: connect with mfsmount, let me list and cd to different folders, then fail to do any file I/O. Additionally, I have realized why my linux client wasn’t working - I had screwed up its network settings.. it was on the wrong network. About the network setup: All 4 systems have Intel X520-DA2 10gbit network cards. There are no firewalls turn on, and no VLANs configured. They’re connected to two switches: - Brocade 10gbit network switch, with 172.31.31.0/24 network, no other traffic here. - Juniper 10gbit network switch, for managing them, this is where my other traffic is, and where I mistakenly put the linux client. Sorry about that, it actually wasn’t until I’d started writing this email regarding the network setup where it finally clicked that the chunk servers are all expecting to only talk on the 172.31.31.0 network, which is *not* where I put that linux client initially. Moving it there solved that issue. I’ll follow this up with some performance testing, and get back to my previous issue, where I feel like the freebsd client performance is poor. Quick dd test on the linux client gives me about a 2.2x performance of what I was seeing on FreeBSD. So, thanks for making me think through my issue again, back to performance testing. -Joe > On Aug 18, 2015, at 1:22 AM, Aleksander Wieliczko <ale...@mo...> wrote: > > Hi Joe! > > First of all I wouldn't call that performance issue because it's look like your client have problem with chunkservers communication. > > Please check your firewall configuration. > Mfsclient use 9422 port to communicate with chunkservers and 9420 to communicate with master. > Does your all moosefs components are in the same network, the same VALN? > > All ports that are used in MooseFS cluster communication are from range 9419-9422 > > I would like to add that default block size for dd is 512 bytes, so you will not achieve good performance not only in MooseFS but in all distributed file systems. > Try to increase block size to something like 4k, 8k, 32k, 64k 1M and see the results. > > Can you tell as something more about your network configurations? > > Best regards > Aleksander Wieliczko > Technical Support Engineer > MooseFS.com <x-msg://30/moosefs.com> > > On 14.08.2015 21:13, Joseph Love wrote: >> Hi, >> >> I was testing out moosefs (again) on FreeBSD, when I ran into what appeared to be a performance issue, which I assumed was related to the client. So, I decided to try a linux client with my FreeBSD master & chunk servers. >> All systems are running 3.0.39, the master & chunk servers are on FreeBSD, and the client is Debian (jessie). >> >> I can mount the moosefs share on the client, it shows me existing content. I can traverse folders, create new folders, ‘touch’ new files, but the second I try and put data into a file, (“echo blah > test.txt” or “dd if=/dev/zero of=test.dd count=1”) the client seems to just lock-up. >> >> It did work with a FreeBSD client, but again, I think there may be a performance issue, as I couldn’t seem to exceed about 140MB/s (SSDs, 10gbit ethernet, many cores, lots of memory) hence why I wanted to try a linux client. Ultimately I don’t care to run linux clients, but I would have liked to see a comparison in performance between the two. >> >> Any ideas why L linux client doesn’t seem to like my FreeBSD master/chunk servers? >> >> -Joe >> >> >> ------------------------------------------------------------------------------ >> _________________________________________ >> moosefs-users mailing list >> moo...@li... <mailto:moo...@li...> >> https://lists.sourceforge.net/lists/listinfo/moosefs-users <https://lists.sourceforge.net/lists/listinfo/moosefs-users> > |