Re: [RFC] nfs: use 2*rsize readahead size

[Wu Fengguang]

On Wed, Feb 24, 2010 at 01:22:15PM +0800, Dave Chinner wrote:
> On Wed, Feb 24, 2010 at 12:18:22PM +0800, Wu Fengguang wrote:
> > On Wed, Feb 24, 2010 at 11:29:34AM +0800, Dave Chinner wrote:
> > > On Wed, Feb 24, 2010 at 10:41:01AM +0800, Wu Fengguang wrote:
> > > > With default rsize=512k and NFS_MAX_READAHEAD=15, the current NFS
> > > > readahead size 512k*15=7680k is too large than necessary for typical
> > > > clients.
> > > > 
> > > > On a e1000e--e1000e connection, I got the following numbers
> > > > 
> > > > 	readahead size		throughput
> > > > 		   16k           35.5 MB/s
> > > > 		   32k           54.3 MB/s
> > > > 		   64k           64.1 MB/s
> > > > 		  128k           70.5 MB/s
> > > > 		  256k           74.6 MB/s
> > > > rsize ==>	  512k           77.4 MB/s
> > > > 		 1024k           85.5 MB/s
> > > > 		 2048k           86.8 MB/s
> > > > 		 4096k           87.9 MB/s
> > > > 		 8192k           89.0 MB/s
> > > > 		16384k           87.7 MB/s
> > > > 
> > > > So it seems that readahead_size=2*rsize (ie. keep two RPC requests in flight)
> > > > can already get near full NFS bandwidth.
> > > > 
> > > > The test script is:
> > > > 
> > > > #!/bin/sh
> > > > 
> > > > file=/mnt/sparse
> > > > BDI=0:15
> > > > 
> > > > for rasize in 16 32 64 128 256 512 1024 2048 4096 8192 16384
> > > > do
> > > > 	echo 3 > /proc/sys/vm/drop_caches
> > > > 	echo $rasize > /sys/devices/virtual/bdi/$BDI/read_ahead_kb
> > > > 	echo readahead_size=${rasize}k
> > > > 	dd if=$file of=/dev/null bs=4k count=1024000
> > > > done
> > > 
> > > That's doing a cached read out of the server cache, right? You
> > 
> > It does not involve disk IO at least. (The sparse file dataset is
> > larger than server cache.)
> 
> It still results in effectively the same thing: very low, consistent
> IO latency.
> 
> Effectively all the test results show is that on a clean, low
> latency, uncongested network an unloaded NFS server that has no IO
> latency, a client only requires one 512k readahead block to hide 90%
> of the server read request latency.  I don't think this is a
> particularly good test to base a new default on, though.
> 
> e.g. What is the result with a smaller rsize? When the server
> actually has to do disk IO?  When multiple clients are reading at
> the same time so the server may not detect accesses as sequential
> and issue readahead? When another client is writing to the server at
> the same time as the read and causing significant read IO latency at
> the server?
> 
> What I'm trying to say is that while I agree with your premise that
> a 7.8MB readahead window is probably far larger than was ever
> intended, I disagree with your methodology and environment for
> selecting a better default value.  The default readahead value needs
> to work well in as many situations as possible, not just in perfect
> 1:1 client/server environment.

Good points. It's imprudent to change a default value based on one
single benchmark. Need to collect more data, which may take time..

> > > might find the results are different if the server has to read the
> > > file from disk. I would expect reads from the server cache not
> > > to require much readahead as there is no IO latency on the server
> > > side for the readahead to hide....
> > 
> > Sure the result will be different when disk IO is involved.
> > In this case I would expect the server admin to setup the optimal
> > readahead size for the disk(s).
> 
> The default should do the right thing when disk IO is involved, as

Agreed.

> almost no-one has an NFS server that doesn't do IO.... ;)

Sure.

> > It sounds silly to have
> > 
> >         client_readahead_size > server_readahead_size
> 
> I don't think it is  - the client readahead has to take into account
> the network latency as well as the server latency. e.g. a network
> with a high bandwidth but high latency is going to need much more
> client side readahead than a high bandwidth, low latency network to
> get the same throughput. Hence it is not uncommon to see larger
> readahead windows on network clients than for local disk access.

Hmm I wonder if I can simulate a high-bandwidth high-latency network
with e1000's RxIntDelay/TxIntDelay parameters..

> Also, the NFS server may not even be able to detect sequential IO
> patterns because of the combined access patterns from the clients,
> and so the only effective readahead might be what the clients
> issue....

Ah yes. Even though the upstream kernel can handle it well, one may
run a pretty old kernel, or other UNIX systems. If it really happens,
the default 512K won't behave too bad, but may well be sub-optimal.

Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/