We have a call open with Sun on this. When they came today to discuss our Platinum systems, we had a chat about this. This is becoming a rather interesting finger pointing problem. Perhaps someone has the knowledge to add value to this.
Solaris 9 has an issue with UFS filesysems over 1 TB due to the EFI label. Apparently, Solaris 10 can go up to about 2 TB as it understands EFI. So, some bright spark in our team broke the 1.4 TB volume into two smaller luns and presented them to the host. VxVM then put them together as a concatenated volume and now we have a VxFS filesystem of about 1.3 TB. So, it seems that all was going ok until recently and the read performance went downhill. This may suggest that concatenation of the two luns has caused the problem. Writes are much quicker (perhaps four times) than reads. I don't know how RAID logical devices are physically made up. We have 11 x 146 GB disks as the logical drive plus a spare. Lets say the logical device has 1.4 TB. I create the first lun using 700 GB and the second lun using 700 GB. I am not using partitioning. When put together, the VxFS file system has about 1.3 TB. So when we write stuff to the filesystem, will it fill up the first part of the concatenated filesystem? In other words, is the first lun written to and when it is full, is the second lun used? If that is the case, then writes are much easier on the device but reads have to scan quite a large area of at least one lun if not both of them. So, if I am on the right track here, I could upgrade the system to solaris 10, present the whole 3510 logical device as one lun, create a non Veratis filesystem with the usedirectio option of about 1.4 TB and our performance problems should go away??? Pretend that all will be ok considering the size of the lun. At least with Solaris 10, I get dtrace and I hope to try to use iosnoop on this. So this leads me to ZFS. How can you troubleshoot ZFS problems? Stephen This message posted from opensolaris.org _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
