Parallelizing calls to msDrawLayer()

[Ed McNierney]

Dave -

If those rasters and shapefiles are on the same disk volume, "parallelizing" them is very likely to actually make things worse!  MapServer will simply get in its own way, causing excessive disk seeks. And that may well be the most common situation for many users.

There are many good ideas here, but not all good ideas work for all scenarios.  Your initial comments seemed focus on rasterization time, not disk I/O optimization, which could be implemented in a rather different manner (one could fetch data in parallel and then render serially, as is done with WMS input).  It's good to bring these ideas out on the table, but this discussion is ranging over a lot of different ideas suitable for a lot of different situations.  It would be helpful to distinguish whether you're trying to (a) provide a built-in optimization for common use cases or a tool that can be used by sophisticated users, and (b) whether you're focusing on drawing speed or data fetch times.

	- Ed

Ed McNierney
Chief Mapmaker
Demand Media / TopoZone.com
73 Princeton Street, Suite 305
North Chelmsford, MA  01863
Phone: 978-251-4242, Fax: 978-251-1396
ed at topozone.com




-----Original Message-----
From: David Fuhry [mailto:dfuhry at cs.kent.edu] 
Sent: Saturday, October 13, 2007 11:19 PM
To: Ed McNierney
Cc: MAPSERVER-DEV at LISTS.UMN.EDU; woodbri at SWOODBRIDGE.COM
Subject: Re: [UMN_MAPSERVER-DEV] Parallelizing calls to msDrawLayer()

Ed,

    Indeed.  I'm very appreciative of all your guys' comments.  I should 
clarify one point.

    I agree that it would not be worthwhile to just parallelize 
rendering.  What I thing might be worthwhile, is to parallelize both 
rendering *and* the I/O that necessarily precedes it.  I view the latter 
as (through no fault of mapserver's) the bottleneck, and the former as 
perhaps icing on the cake.  Since the late bird doesn't get the worm ;), 
Ed, may I use you for an example?

    You serve lots of raster.  Let's say your mapfile is composed of a 
base raster layer and a few vector layers on top.  A request is made. 
mapserv goes to render the first layer.  The tileindex is probably in 
the page buffer (memory), so it looks up the tile(s) quick and goes to 
fetch the raster image(s).

    Maybe the images have to come across NFS.  The request goes over 
GigEth; ping / 2 says this takes 36ms.  The fileserver seeks its 7200RPM 
disk to the start of the TIFF; Seagate says this takes 8.5ms.  Let's say 
the sequential read & transfer back take zero time, except the 36ms 
lower-bound on network time.

    What has mapserv done in this 36 + 8.5 + 36 = 80.5ms?  Nothing. 
Just waited on I/O.  It could have perhaps rendered several vector 
layers from ramdisk in this time.

    Now Frank's GDAL goes to work mosaicing/clipping/warping/resizing 
the raster image(s).  It's chewing CPU.  Let's say that layers 2 thru n 
are shapefiles on the local disk.  What is the local disk doing during 
this time?  Nothing.  It could be checking for the existence of 
layer2.qix, or seeking to the start of layer2.shp, or likewise for 
layers3 thru n.  None of which will interfere with GDAL's work. 
Instead, we wait until GDAL is through, then waste 8.5ms seeking to the 
start of layer2.shp.  We could have sought layer2.shp even earlier, 
while waiting on the NFS request.

    The thing which seems beautiful to me, is that OS schedulers (both 
process & I/O) are designed to be good at receiving a bunch of requests, 
and resolving them efficiently.  I think there may be value in launching 
a thread for each layer, thus throwing all the requests up against the 
OS at once, and letting its schedulers try to make the best use of CPU 
and I/O resources.  I have to imagine that they are likely to do better 
than having everything wait in line.

    Yes, this would put more of a strain on the server at that instant. 
  By pipelining I/O though, it will also return a map quicker.

Thanks,

Dave


Ed McNierney wrote:
> David -
> 
> I didn't have the time for a thoughtful reply earlier, and now most other folks have already raised some of the concerns I had.  I should hesitate more often, I guess - it saves typing <g>.
> 
> I think Steve Woodbridge's comment is informative.  For his application he found a 4x-5x improvement by caching data files in a RAM disk.  That basically says that something like 80% of his entire MapServer rendering time is spent in disk I/O, not drawing.  Many users don't have a situation in which they can put their data in RAM disk.  For a comparable kind of application you would then reasonably predict that optimizing multi-layer rendering so it was instantaneous would only produce a 20% performance improvement.
> 
> Although I think MapServer's disk I/O is pretty good, if I were to spend time hunting for performance improvements I would be inclined to look at the various data I/O schemes.  Anything that can be done to reduce disk I/O is a big win (some of those improvements are, of course, external to MapServer itself in the form of data organization and indexing schemes).
> 
> 	- Ed
> 
> Ed McNierney
> Chief Mapmaker
> Demand Media / TopoZone.com
> 73 Princeton Street, Suite 305
> North Chelmsford, MA  01863
> Phone: 978-251-4242, Fax: 978-251-1396
> ed at topozone.com
> 
> 
> 
> -----Original Message-----
> From: UMN MapServer Developers List [mailto:MAPSERVER-DEV at LISTS.UMN.EDU] On Behalf Of David Fuhry
> Sent: Saturday, October 13, 2007 8:39 PM
> To: MAPSERVER-DEV at LISTS.UMN.EDU
> Subject: Re: [UMN_MAPSERVER-DEV] Parallelizing calls to msDrawLayer()
> 
> Paul,
> 
>     Thanks, that's a good suggestion.
> 
>     I guess my thought is, given a really good implementation, a 
> heavily-contended server with a bright scheduler would just end up 
> scheduling the threads sequentially on the same CPU (perhaps likely, 
> since a small bit of the necessary data is in that processor's L1 cache 
> already).  Then the onus is on the implementer to make sure that the 
> extra overhead is pretty low.
> 
>     It sort of pushes some of the responsibility to the OS scheduler. 
> Which I think most of the time, will make better decisions than will a 
> deterministically-ordered mapserv loop.
> 
> Thanks,
> 
> Dave
> 
> Paul Spencer wrote:
>> David,
>>
>> While you can perhaps gain some performance in a single map draw, in 
>> most real life uses of mapserver, folks are either serving many 
>> simultaneous requests or generating tiles in some way.  I think in 
>> either case, the addition of multi-threaded layer draws will actually 
>> cause contention for processor time with the multiple processes that are 
>> serving the requests and could hurt overall performance in high load 
>> systems.
>>
>> I think that you could probably get more bang for your development bucks 
>> by investing time in profiling the existing code.
>>
>> Cheers
>>
>> Paul
>>
>> On 13-Oct-07, at 6:37 PM, David Fuhry wrote:
>>
>>> Tamas,
>>>
>>>    (responses inline)
>>>
>>> Tamas Szekeres wrote:
>>>> David,
>>>> I consider it would be reasonable to establish such mechanism only
>>>> when fetching the data of the layers. Likewise currently the WMS/WFS
>>>> layers are pre-downloaded in parallel before starting to draw the map.
>>>> We should have a similar approach when fetching the other layers as
>>>> well.
>>>    Yes, I noticed that WMS/WFS layers are downloaded in parallel 
>>> before rendering begins.  And I agree, it would be advantageous to 
>>> extend the parallel-data-fetching paradigm to all layers.
>>>
>>>    For non-WMS/WFS layers though, wouldn't it be a significant 
>>> disruption to the codebase to add lines 1 and 2 into msDrawMap()?
>>>
>>> 1. for i=1 to layers.length (in parallel)
>>> 2.   data[i] = fetch_data_for_layer(i)
>>> 3. for i=1 to layers.length (serially)
>>> 4.   msDrawLayer(data[i])
>>>
>>>   ISTM that the data-fetching logic might be best left abstracted 
>>> beneath msDrawLayer().
>>>
>>>> However pre drawing all of the layers and later copying the layers
>>>> over the map image seems to be much less efficient.
>>> Drawing n layers onto n imageObjs is no more expensive than drawing n 
>>> layers onto one imageObj, and the former can be parallelized across n 
>>> threads.
>>> Although yes, I agree that composition (the "merge" step) will cost 
>>> something.
>>> I'm entertaining the idea that the time saved by parallel fetching & 
>>> drawing might outweigh the cost of composition.
>>>
>>>> When using the parallel fetching approach we should deal only with the
>>>> drivers from the aspect of the thread safety issues.
>>> I might be misunderstanding your point here, but... Rendering a layer 
>>> into an independent imageObj should be a pretty independent operation, 
>>> and could be made so if it's not now.  Glancing at the mapserver 
>>> thread-safety FAQ, it seems there are more unsafe & locked components 
>>> related to data-fetching drivers than there are for rendering.  Which 
>>> makes me wonder why you suggest parallelizing the data-fetching but 
>>> not the rendering.
>>>
>>> Forgive me if I'm playing a bit of devil's advocate here.  I'm aware 
>>> that non-reentrant functions don't rewrite themselves, and that 
>>> critical sections don't surround themselves with mutexes.  Surely 
>>> though, it ought not to be a tremendous amount amount of work to keep 
>>> separate layer-drawing operations from stepping on eachothers' toes?
>>>
>>> Thanks,
>>>
>>> Dave Fuhry
>>>
>>>> Best regards,
>>>> Tamas
>>>> 2007/10/12, David Fuhry <dfuhry at cs.kent.edu>:
>>>>> Has anyone looked into parallelizing the calls to msDraw[Query]Layer()
>>>>> in msDrawMap()?
>>>>>
>>>>> Although I'm new to the codebase, it seems that near the top of
>>>>> msDrawMap(), we could launch a thread for each (non-WMS/WFS) layer,
>>>>> rendering the layer's output onto its own imageObj.  Then where we now
>>>>> call msDraw[Query]Layer, wait for thread i to complete, and compose 
>>>>> that
>>>>> layer's imageObj onto the map's imageObj.
>>>>>
>>>>> In msDraw[Query]Layer(), critical sections of the mapObj (adding labels
>>>>> to the label cache, for instance) would need to be protected by a 
>>>>> mutex.
>>>>>
>>>>> A threaded approach would let some layers get drawn while others are
>>>>> waiting on I/O or for query results, instead of the current serial
>>>>> approach where each layer is drawn in turn.  Multiprocessor machines
>>>>> could schedule the threads across all of their cores for simultaneous
>>>>> layer rendering.
>>>>>
>>>>> It seems this could significantly speed up common-case rendering,
>>>>> especially on big machines, for very little overhead.  Has there been
>>>>> previous work in this area, or are any major drawbacks evident?
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Dave Fuhry
>>>>>
>> +-----------------------------------------------------------------+
>> |Paul Spencer                          pspencer at dmsolutions.ca    |
>> +-----------------------------------------------------------------+
>> |Chief Technology Officer                                         |
>> |DM Solutions Group Inc                http://www.dmsolutions.ca/ |
>> +-----------------------------------------------------------------+