Richard McDougall's Weblog
Richard McDougall's Weblog
Commentary from Race Control

20051111 Friday November 11, 2005

Tuning for Maximum Sequential I/O Bandwidth

John asks the question "what does maxphys do, and how should I tune it?"

The maxphys parameter use to be the authoritive limit to the maximum I/O transfer size in Solaris. A large transfer size, if the device and I/O layer supports it, generally provides better large I/O throughput. This is generally supported by the fact that disks like larger transfers if you are trying to get absolute maximum throughput. For a single disk, 64k is generally the point at which maximum transfer rate occurs, and for disk arrays, typically 1MB.

Historically, there was a maxphys of 56k set, due to some older (VME?) bus max xfer limitations. The maxphys limit was increased to 128k on SPARC around the Solaris 2.6 timeframe. Since Solaris 7, the sd/ssd drivers (SCSI and Fibre Channel) override maxphys if the device supports tag queuing, up to a default of 1MB.

On x86/x64, the default is still 56k (we need to look at this!).

In summary, the defaults for SPARC with SCSI or Fibre channel are optimal defaults. You can always check by doing a large I/O test and observing the average I/O size:

# dd if=/dev/rdsk/c0t0d0s0 of=/dev/null bs=8192k

# iostat -xnc 3
     cpu
 us sy wt id
 10  4  0 86
                    extended device statistics              
    r/s    w/s   kr/s   kw/s wait actv wsvc_t asvc_t  %w  %b device
   22.4    0.0 22910.7    0.0  0.0  1.0    0.1   44.6   0 100 c0t0d0

Here we can see that we are doing 22 reads per second, and 22MB/s; thus the disks are performing optimal 1MB I/O's. Also, a simple set of test with varying block sizes will help identify the best I/O size for your device.

One more comment about max transfer size; these parameters (the driver's max transfer size) is read by UFS's newfs command, and used to set the cluster size of the file system; i.e. the number of contiguous blocks to read ahead or write behind. If you put a file system on a device and want to see large I/O's, you'll need to ensure that the maxcontig parameter also reflects the devices's max transfer size. It can be tweaked after with tunefs.

# mkfs -m /dev/dsk/c0t0d0s0
mkfs -F ufs -o nsect=248,ntrack=19,bsize=8192,fragsize=1024,cgsize=22,free=1,rps=120,nbpi=8155,opt=
t,apc=0,gap=0,nrpos=8,maxcontig=128,mtb=n /dev/dsk/c0t0d0s0 18588840

Here you can see that maxcontig=128 8k blocks, which is 1MB. If you tune maxphys, then reset the file system's max cluster size with tunefs after, too.

[ T: ]

( Nov 11 2005, 12:12:55 PM PST ) Permalink Comments [9]



Archives
Referrers