As mentioned here I got my new server working with a 3ware 8006-2LP and a pair of new 500GB disks. While it was working fine, I noticed that when updating the FreeBSD ports collection that the update would occasionally pause, consuming no CPU, but with the update process having a status of sbwait. I understand this to mean that the process is waiting on a blocked socket.
It turns out that the twe(4) driver is what is known as GIANT-LOCKED, which I believe means that it uses the old SMP locking mechanism in FreeBSD:
twe0: <3ware Storage Controller. Driver version 1.50.01.002> port 0x8c00-0x8c0f mem 0xfc7ffc00-0xfc7ffc0f,0xfb800000-0xfbffffff irq 28 at device 3.0 on pci1
twe0: [GIANT-LOCKED]
twe0: [ITHREAD]
twe0: 2 ports, Firmware FE8S 1.05.00.068, BIOS BE7X 1.08.00.048
Best I can tell, the result of this is that the disk controller’s driver needs to wait for the kernel to free up other resources and tell the driver that it can go ahead and work before it does things. The result of this tends to be that the driver works well, but there is a lot of latency.
This understanding matches what I observed, which was the aforementioned lengthy pauses when doing things which required a bunch of disk IO. In order to prove this understanding out, I set up a test hard disk running a stock FreeBSD 7.0 amd64 installation from which I could run Bonnie++, a file-based disk benchmarking suite.
In my testing I used the following three scenarios:
· One 120GB IBM Deskstar PATA drive (IC35L120AVVA07) connected to the motherboard booting the OS, listed in the results as banstyle_deskstar.
· Two 500GB Western Digital SATA drives (WD5000AAKS-40TMA0) connected to the motherboard with software RAID 1 via gmirror(8), listed in the results as banstyle_gmirror.
· Two 500GB Seagate SATA drives (ST3500320AS) connected to the 3ware 8006-2LP using the twe(4) driver in hardware RAID 1, listed in the results as banstyle_twe.
The result ended up being that all three configurations are generally around the same speed for throughput, but the 3ware controller had an absurd amount of latency. If one looks at the HTML version of the Bonnie++ output here (or the PNG here or above), one can see that was giving near three SECONDS of latency for random seeks and writes using write(2). This is insane.
The only thing I can think to attribute this to is the GIANT-LOCK in twe(4). I guess this means that I’m going to have to go back to gmirror(8) for software RAID and return the card. How disappointing.
(If anyone reading this disagrees with these findings or wishes to comment on them, please don’t hesitate to do so here or by emailing me directly.)
7 Comments