Conference decwet::winnt-clusters

    
    
    They said they would have a DLM for WOLFPACK?   
    
    Hmmmm.  The last presentation I heard was they had done an industry 
    analysis and determined there was no need for a DLM.  They indicated
    that with 'shared nothing' DLM is not necessary...  
    
    Mary Jo 

    I agree with Mary.. that was what they said in the last presentation I
    attended...
    They also mentioned that software vendors might build DLM on their own..
    (Oracle is doing so for their Oracle Parallel Server product, which
    runs on top of WOLPACK...)
            
    - Murali

    I double checked my info with the Microsoft presentor and he claims
    there will be a DLM in the first release of Wolfpack.  I was very
    surprised to hear this myself.
    
    BTW, Compaq and Tandem also did short presentations on their cluster
    solutions.  Compaq stressed low cost and ease of deployment.  They will
    have something called an Integrator which supposedly goes out and
    probes the hardware to make sure it's supported on their cluster and at
    the right rev levels.  They will also have fibre channel by the end of
    summer '97.
    
    Tandem talked about several interesting contributions to clustering. 
    They briefly mentioned "heartbeat" technology which just sounded like
    "I'm alive" polling.  There was also discussion of ServerWare which
    automatically clones executables to be run on each processor as load
    demands.  ServerWare is supposed to ship in a year.  The most
    interesting technology from Tandem is ServerNet.  With ServerNet Tandem
    claims a 300 mbyte/sec throughput with a 300 ns latency.  The ServerNet
    architecture will be implemented on a single chip.  The presentor also
    claims that ServerNet will be able to achieve 1.2gbyte/sec throughput
    "soon" after initial release.  Tandem will also ship a fault tolerant 4
    processor 200 mhz Pentium Pro server based on their S1000RM server. 
    
    I would like to know how much of all of this (ie Microsoft, Compaq and
    Tandem) is marketeering and how much is reality.
    
    
    gary
    

    We have talked to the Microsoft Wolfpack program and product managers,
    and both say that Wolfpack V1.0 will not ship a DLM.  At this point,
    their attitude is that someone will have to present a very good
    business case for WHY someone would want a DLM.  (What applications
    would use it?)
    
>    Compaq stressed low cost and ease of deployment.  They will
>    have something called an Integrator which supposedly goes out and
>    probes the hardware to make sure it's supported on their cluster and at
>    the right rev levels.
    
    We already have a similar tool for Digital Clusters called "cluivp"
    (CLUster Installation Verification Procedure). 
    We will probably ship a version of that for wolfpack too.
    
>    With ServerNet Tandem
>    claims a 300 mbyte/sec throughput with a 300 ns latency.  The ServerNet
>    architecture will be implemented on a single chip.
                 ^^^^^^^
    
    Notice that phrase "will be".  Not quite ready yet, but it does look
    promising.  Digital's Memory Channel provides similar throughput and
    latency.  Memory Channel has been used by Digital's UNIX and VMS
    clusters for over a year now, and we're working on drivers for NT right
    now.

    Hi Carl,
    
    Thanks for your reply.  It's interesting how much message distortion
    occurs between corporate headquarters of any organization, in this case
    Microsoft, and the field.  I'll make sure my local Microsoft contact is
    aware of the discrepancy and I'll suggest he verify his DLM claim. 
    Could you perhaps give me the name of the Microsoft cluster people that
    he should contact?  This is a Microsoft corporate cluster pitch that
    this guy is presenting from so Microsoft corporate will have to issue a
    correction to all their field offices I would assume.
    
    Regards,
    gary
    

    The Product Manager for Wolfpack at Microsoft is Mark Wood.
    
    Early in the product cycle (August 1996), MS said they WOULD ship a DLM
    with Wolfpack V1.0.  Now they have changed their message to say they
    will NOT ship a DLM.
    
    If you have customers who express a need for a DLM, can you find out
    why they need one?  What would they use it for?

>    Early in the product cycle (August 1996), MS said they WOULD ship a DLM
>    with Wolfpack V1.0.  Now they have changed their message to say they
>    will NOT ship a DLM.
>    
>    If you have customers who express a need for a DLM, can you find out
>    why they need one?  What would they use it for?

Presumably so that the servers could access all disks simultaneously just like
a VMScluster. Far better for load balancing and simpler for management too as
there would be no need or concept of failing over a disk from one server to
another as it would already be accessible.

A better question might be. "What customers don't need a DLM?"

Nigel

    
    The following is a description of shared nothing -- a reason why 
    Microsoft doesn't consider DLM a requirement. 
    
    Remember that load balancing is not intended to be a feature until 
    a later phase of Wolfpack.  
    
Shared Nothing Model

In the shared nothing software model, each system within the cluster owns a
subset of the resources of the cluster. Only one system may own and access a
particular resource at a time, although, on a failure, another dynamically
determined system may take ownership of the resource.  In addition, requests
from clients are automatically routed to the system that owns the resource.  

For example, if a client request requires access to resources owned by multiple
systems, one system is chosen to host the request.  The host system analyzes
the client request and ships sub-requests to the appropriate systems.  Each
system executes the sub-request and returns only the required response to the
host system.  The host system assembles a final response and sends it to the
client. 

A single system request on the host system describes a high-level function
(e.g., a multiple data record retrieve) that generates a great deal of system
activity (e.g., multiple disk reads) and the associated traffic does not appear
on the cluster interconnect until the final desired data is found.  By
utilizing an application that is distributed over multiple clustered systems,
such as a database, overall system performance is not limited by a single
machines hardware limitations. 

The shared disk and shared nothing models can be supported within the same
cluster.   Some software can most easily exploit the capabilities of the
cluster through the shared disk model.  This software includes applications and
services which require only modest (and read intensive) shared access to data
as well as applications or workloads that are very difficult to partition. 
Applications which require maximum scalability should use the clusters shared
nothing support.


    

>Presumably so that the servers could access all disks simultaneously just like
>a VMScluster. Far better for load balancing and simpler for management too as
>there would be no need or concept of failing over a disk from one server to
>another as it would already be accessible.
>A better question might be. "What customers don't need a DLM?"

Let's not confuse the disk access mechanism from the data synchronization 
mechanism, although they are related.

For the disk access models, there is the shared approach where the disk is 
phyically connected the systems via a storage bus, such as Scsi for Fibre 
Channel.  There is also the shared nothing approach, where the disk is only 
connected to one or a small subset of the nodes, and the disk blocks are served 
out to the remaining nodes.  Data access performance is important in shared 
nothing systems so a high speed bus of some kind may be used, such as FDDI or 
Memory Channel, to ship the blocks around.  The benefit of shared nothing in my 
mind is that generic hardware (servers, network interconnects) scales better 
and less expensively than dedicated storage buses (disk controller subystems,
storage interconects) at some cost in performance. 

In two models are equivalent, just implemented at different levels.  A server 
provides blocks to clients over an interconnect.  The server may be redundant 
and may support raid algorithms.

Assuming that low level data sharing is available, what is the best way to 
synchonize access to those blocks?  A DLM is certainly one way to appoach this. 
Partitioning the data might be another.  Changing the kind of data being 
distributed might simplify the locking requirements depending on the
application.

Perhaps we need to rephase the question.  Is a shared data model cluster where 
all noes can access the storage better than a two node failover based cluster?  
The answer is yes, assuming you want to change your filesystem, database and 
other disk intensive applications.

Is a user-visible DLM the best model for distributed applications? Debatable.  
How many applications maintain disk block caches? Parallelizing a database or
a filesystem is a major effort. Such a thing is difficult to implement,
maintain and prove correct.  Performance is also an elusive thing since poor
lock design will result in lock contention.   For some applications, a DLM is 
the right thing.  Oracle Parallel Server comes with its own.  What other 
application needs one?  Distributed system toolkits such as ISIS may be a better
choice for general application coarse grained parallel programming.

Is the question, why shouldn't we make NT clusters be just like VMSclusters?  I 
think the answer is that NT does most of what VMS did in a cheaper, simpler and 
more extensible way.  There is no doubt that VMSclusters was state of the art
for its time, but it also required propriety hardware, did not scale well, and 
was very intertwined with the lowest level operation of the system.  True, NT 
does not have a shared system disk, but it has domains for a common 
administraton environment and LAN-based protocols for file sharing.  Loosely 
coupled client-server architecture is much easier to extend than tightly
coupled low-level synchronization protocols.

    
    VMSClusters are great.  However, when we start talking terabyte and 
    pentabyte databases in a shared disk environment, the overhead from
    disk io in a cluster could be overwhelming. 

   ho hum.
    
    Microsoft have all their heaviweights lined up against OpenVMS-type
    filesystem's.  Cutler (who left VMS before V4) goes (sic) 'nonlinear'
    when they are discussed.
    
    Jim Gray who has a huge credibility in the field of databases is
    proseletizing for shared nothing.  For instance, when I challenged him
    about Oracle's OPS (the flagship shared everything model) he asserted
    that they actually prefer a explicit load balancing (eg shared nothing)
    mechanism for their TPC results.  That week, someone else from
    micrsoft told me that the CTO of Oracle was a believer in shared
    nothing (I have not validated that, I just offer it as an example of
    the hearts and minds offensive)    
    
    Technically, there are real concerns in the NT kernel group about the
    interaction of NT's Cachemanager and a shared everything filesystem. 
    Indeed, the sharing mechanism needs to be very carefully thought through
    so that you don't thrash your caches.  I have to admit that the
    Wolfpack Cluster people know their stuff and have really examined and
    thought through what clusters actually mean.
    
    So, politically and in marketing terms a shared everything environment
    is hard. Who has the better marketing, DIGITAL or Microsoft ?
    
    Now, Personally speaking I think that there is a place for shared
    everything, between the ultimate top end cluster and SMP.  The
    important characteristic seems to be the immediate transportability of
    an environment from Node to Node (dynamic reconfiguration) and the
    flexibility that this gives the system manager.
    
    However, in an environment where every penny of expenditure has to be
    carefully examined.  If Digital does this, what will not get done. 
    This is why way back in .-n, Carl is asking for which customers would
    be able to take advantage of DLMs and of shared everything storage.
    
    So, this is not about the technology, but about the market.
    You, the people reading this, you are the one with the customer
    contacts and information which will or will not drive this forward. 
    
    If you have information on the commercial potential of this, I urge you
    to contact Carl..
         

	I think, too, that much of the wailing over the lack of shared
	everything comes from the VMScluster-knowledgeable people because
	that is what they know.  I have worked with all 4 clustering
	technologies from Digital (TOPS-20, VMS, UNIX, NT).  I had to be
	persuaded because I understood the shared everything and the
	shared nothing was new.  But having talked to hundreds of
	customers about everything vs. nothing, about the only people who
	bring up the idea of shared everything are the VMS people.  Then
	I ask them how they design arrange their data for maximum performance
	in a VMScluster.  Except for rare cases, they always assign the
	data to a non-shared disk.  In other words, they make it shared 
	nothing so they don't have to deal with the overhead of DLM.

	There are places for both; I'm not arguing that.  I think, though
	that the shared nothing model is being lambasted too quickly.  To
	expect V1 products to be as robust as VMSclusters is a little too
	much to expect.  Shared nothing was the easiest way to introduce
	clusters to NT in a manner that is very effective for what it is
	trying to do - providing high-availability to data and services.
	If the market (other than just former VMScluster people) sees a
	need for shared everything, I am sure MS will do something about it.

tgc

    I hope that everyone contributing to this discussion also understands
    that we're talking about 2 separate things:
         
         A Distributed Lock Manager, which is one way to enable cooperating
         applications among nodes in a cluster.
         
         A shared filesystem, where several nodes in a cluster access
         common bits on common disks, and somehow coordinate access among
         themselves.  (One way to coordinate access is to use a Distributed
         Lock Manager.)
    
    I think that eventually, you'll see Microsoft produce an API that
    allows applications on cluster nodes to cooperate among themselves. 
    I doubt they'll do a shared filesystem, however.

Gary:

I exchanged email with Tom Phillips the Wolfpack Product Manager and you
are correct we will not be supporting DLM in the first release of
Wolfpack and not in any future version as well. We will be removing in
references to DLM from our presentations. 

Dennis 

    
    
    I just love being right.
    
    Mary Jo 
    

    re .15
    
    Have you ever been wrong?

    
    I thought so once, but I was mistaken...     ;-}
    
    Mary Jo 

    >I thought so once, but I was mistaken...     ;-}
    
    Good thing we don't have to have an AI program
    analyze THAT one. /m

    
    AI.  Don't get me started on AI.  Having to ship those containers
    by FEDEX to make sure they get there overnight...
    
    Mary Jo 
    

	I always had problems with the acronym AI.  I was raised on a
	farm.  AI (the farmers will know) has a significantly different
	meaning.  The Artificial is the same, but the 'I' part never made
	any sense, until I started hearing about 'hybrid' computers.

tgc







p.s.  for the non-farmers - AI is artificial insemination.

    
    Perhaps someone can clarify something for me:
    
    Wolf Pack addesses availability (2 systems) .. no problem.
    
    Now, given a single large database with a cpu intensive application.
    With Wolf Pack and no DLM, all users access the db via a single node.
    Now after a few months, the cpu gets peaked out (say a quad P6) 
    
    ..what are the options?
    
    How does one address cpu scalability in a Wolf Pack environment ?
    
    In an OpenVMS environment with DLM, one simply drops a new system into
    the cluster and load sharing begins ie. cpu resources are shared. You
    do not even need to reboot any other systems.
    
    Am I missing something here ?
    
    Thx in advance,
    
    / Kerry
    
    

    re: .24 by OTOU01::MAIN
    
    >> What about cpu scalability ?
    >> Am I missing something here ?
    
    Are these rhetorical questions?  ;-)

Rod,

Thx for the info - I have a few additional issues which hopefully either
yourself or someone else here can expand on for me ..

This whole debate seems to come down to "does a DLM add value to both
the DB and timesharing (thin client/fat server) environments or just
the DB environment or neither ?

I would suggest both, but for the purposes of this note, will only address 
the DB environment.

>>>    
    The conceptial difference between shared everything and shared nothing
    when applied to scaling is one of data-division as opposed to
    algorithmic division.  The argument is that all data is susceptable to
    data division, so as your database scales to a terabyte and beyond you
    apply an a-priori split and use a TP monitor (& Viper) to channel the
    job to the processor needed.  VI/SAM will serve to pass messages
    between nodes which are involved in a custom transaction.

    ....

    Shared nothing means that caching is easier (especially a write-back
    environment), is easier to program to and has less communication costs
    than shared everything.  But it does rely on up front static data
    division to work.
>>>

I am having a really hard time trying to understand why Microsoft says
they dont need a DLM, yet Oracle is building a DLM into their new parallel 
server. Is Microsoft really not planning a DLM for SQL Server ?

Note - arguably Oracle has many of the best DB type folks (who understand 
real production environments) in the world - not Microsoft, who only recently 
added row level locking to SQL Server..

It appears that the marketing folks from Microsoft are trying to convince 
the world they know production db systems better than anyone else, but, in an
attemp to keep a somewhat open mind, perhaps someone could expand on the 
following concerns that I have with the "shared nothing" cluster:

1. Very little flexibility to adapt to reality of changing db work loads.

As you stated, "But, it does rely on on up front static data division 
to work". 

Now, how many real life Cust db's or work loads remain static ? 

New tables, indexes are constantly being added, modified, dropped etc, 
depending on the changing requirements of their business. What works well in 
the lab, at installation and in theory are usually totally different from what 
happens when a ton of "no mercy" users are thrown into the equation. 
    
Soooo, if the workload changes, and a single node becomes a bottleneck, does 
this mean a cluster reconfig with down time ? Does it mean a db reorg ? The 
DB queries are at the mercy of the internal db optimizer, so does this not 
further complicate the real life performance expectations  and trying to 
assign specific nodes to specific resources?

Bottom line - How will the shared nothing cluster be adapted to changing 
workloads? 

2. Load balancing is not part of Wolf Pack phase 1. Fine, but that in effect
   is saying that it is not scalable.  Has Microsoft stated how they will do
   future load balancing (add scalability) without a DLM ?

   To expand, if a single cpu is designated to handle one type of workload, 
   what happens when that one workload exceeds the CPU capacity of a single 
   system? 

   Buy a bigger system, ok, but what happens in an Intel Cluster when a quad 
   system gets maxed out on cpu resources - very easy to do with complex db 
   queries and/or thin client / fat server environment ?

   Adding more systems means nothing - you must rewrite or adapt the 
   application ! Just ask any Customer which would be easier - rewrite an
   application (especially with the heat on) or just drop another system
   into the cluster. Remember that while large prod environments are usually 
   concerned primarily (not exclusively) with performance than cost ... 

   SW development and changes are big time cost (time and $'s), while HW is 
   cheap - even with proprietary solutions. $10k are rounding errors in these 
   environments ...

   Also, keep in mind that while there is overhead in a a shared everything
   environment, it typically is only a few % points of the overall load and 
   the systems we have today are hugely more powerful than in days gone by.

3. Does the shared nothing cluster consider the implications of VLM (very
   large memory) DB's and high speed memory interconnects such as memory 
   channel ? It appears that the shared nothing approach is designed to solve 
   VLDB (big db's on disk) only. Big difference between VLM and VLDB.

   Using a DLM and VLM db (what Oracle is doing) and being able to directly 
   address other systems memory, does this not seem a much more scalable 
   db solution than shared nothing approaches?

As a general statement, could it be stated :

For the DB environment, the shared nothing proponents assume that Customers 
have a good understanding of their workload environments and how they will 
change. It assumes Customers have the expertise to change their application
should a single CPU become a bottleneck. For these environments, shared
nothing clusters are a good fit.

The Oracle and DLM approach is that the workload will be shared across all 
systems - regardless of how the business workload changes.

[I think Oracle are licking their chops right now ..]

:-)

Thx in advance,

Regards,

/ Kerry
    
    

>     
>     I urge all DIGITAL people who are interested to:
>     	1) Keep an open mind and
>     	2) Read Greg Pfeister's book

    Certainly interested -- what is this book?
    
    ed

.24 says:
    
>    Now after a few months, the cpu gets peaked out (say a quad P6) 
>    
>    ..what are the options?
    
    Easy, drop in a quad AlphaServer 4100 instead.  That's what Microsoft
    datacenter is thinking about.  Good for us!
    
>    In an OpenVMS environment with DLM, one simply drops a new system into
>    the cluster and load sharing begins ie. cpu resources are shared.
    
    True for a timesharing load.  How does it work with a client/server
    database setup?  (NT's typical use.)  Can you simply drop a CPU into a
    VMS cluster and have it start processing DB requests?  What software is
    used?  (RDB? Oracle?).  The point is that VMS DB applications like RDB
    and Oracle have been taught to use DLM to coordinate access to shared
    data.  There aren't any such NT applications (yet).  SO at the moment,
    it's sort of a chicken and egg situation.  No DLM -> no applications to
    take advantage of it.  No DLM-enabled applications -> no need for DLM.
    
    We in NT Cluster engineering continue to consider the possibilities of
    DLM.  But it's not like we can write one in a week.  (And the VMS one
    is too "vms-like" to port easily.)

In reply to .28:

In Search of Clusters - The coming battle in lowly parallel computing
by Gregory F. Pfister

published by Prentice-Hall
ISBN: 0-13-437625-0


and on the cover...

'It was originally conceived as a cluster of dogs battling a Savage 
Multiheaded Pooch and, of course, winning.  However, it could also 
represent a system administrator's view of many clusters...'

sound familiar ??

>I am having a really hard time trying to understand why Microsoft says
>they dont need a DLM, yet Oracle is building a DLM into their new parallel 
>server. 
    
    At a top level, by keeping all the information within on processor, all
    conflict of access is kept within that processor.  Hence any `lock
    management' is non distributed.  So a (realtively) heaviweight
    distributed lock is replaced by a loghtweight semaphore/mutex/whatnot
    
>    Is Microsoft really not planning a DLM for SQL Server ?
    
    I would be very suprised.
    
>   Adding more systems means nothing - you must rewrite or adapt the 
>   application ! Just ask any Customer which would be easier - rewrite an
>   application (especially with the heat on) or just drop another system
>   into the cluster. Remember that while large prod environments are usually 
>   concerned primarily (not exclusively) with performance than cost ... 
    
    This is a two edged sword.  In order to take advantage of `just
    slotting another node in', your application needs to have been written
    in a shared everything friendly fashion.  If you are just taking an
    application which runs on one node, it is (conceptually) easier to just
    split the data up.
    
    Of course Microsfot have historically been happy to make application
    writers change their code and seem enthusiatic about making people
    write Viper in Visual basic...
    
    Finally, VLM.  Interetsing arguements to be made again on both sides. 
    But think that VLM tends to be addressable by a single node, and hence
    is intrinsically shared nothing. 
    
    When memory becomes sharable between processors we usually think of SMP
    as being the solution, not clusters..
    
    Just some thoughts. 
    	rod

    re: .29 -
    
    >>>
    Easy, drop in a quad AlphaServer 4100 instead.  That's what Microsoft
    datacenter is thinking about.  Good for us!
    >>>
    
    Not possible in an Intel based Cluster. You would have to replace all
    nodes. Of course that is ok to, assuming Cust wants to do this, simply
    because one node is overloaded cpu wise.
    
    The point remains that the application would have to be re-written or 
    db reorganized should a single SMP node resources be exceeded. Given
    large complex queries in data warehousing environment, topping out a 
    single Intel box will be relatively easy to do. 
    
    >>>
    True for a timesharing load.  How does it work with a client/server
    database setup?  (NT's typical use.)  
    >>
    
    Microsoft is rapidly moving towards adopting thin client - fat server
    environments. This is not unlike the timesharing of the old days.
    Citrix has gained a lot of attention recently by allowing thin clients
    to run app's ON THE SERVER and only send display updates to client.
    
    >>>
    Can you simply drop a CPU into a VMS cluster and have it start processing 
    DB requests?  What software is used?  (RDB? Oracle?).  
    >>>
    
    Yes to both. Numerous other app's also plug-n-play in this environment.
    
    >>> The point is that VMS DB applications like RDB and Oracle have been 
    taught to use DLM to coordinate access to shared data.  There aren't any 
    such NT applications (yet).  SO at the moment,it's sort of a chicken and 
    egg situation.  No DLM -> no applications to take advantage of it.  
    >>>
    
    Bottom line is in the delivery of performance and availability. You can
    bet Oracle's marketing will be promoting large DB VLM numbers very soon
    after NT5 VLM Beta becomes available (even if informal) .. they will
    also likely be pointing out the advantages of their DLM.
    
    How will Wolf Pack scale in thin client / fat server cluster
    environment ? Answer - tbd .. you need to spread load across all
    systems and there is no database to partition, only a traditional
    timesharing load to distribute across all nodes - all of which can't
    be allowed to write over each other, so ?
    
    >>>
    No DLM-enabled applications -> no need for DLM.
    >>>
    
    Most LAN and UNIX based customer's and vendors do not understand the 
    implications of clusters yet, so their simple answer is why? Just
    mention the word "data corruption" and see how fast they start to
    listen.. :-)
    
    The big guy's who have been there, done that, simply smile and continue
    on their merry way ..
    
    :-)
    
    Should be interesting .. especially as thin client, fat server really
    starts to heat up ..
    
    Regards,
    
    / Kerry

    re: .31 -
    
    >>>
    Finally, VLM.  Interetsing arguements to be made again on both sides. 
    But think that VLM tends to be addressable by a single node, and hence
    is intrinsically shared nothing. 
    >>>
    
    Big data warehousing applic's with real scalability issues to consider
    are using multiple VLM nodes and using memory channel technologies to
    directly access each others physical memory. This not only speeds up
    complex db queries, but also drastically reduces amount of expensive IO
    required to do lock management.
    
    At least that's my understanding of what the big guys are up to .. I
    suspect (not sure), that this is how Oracle is planning to do their
    100,000 TPCM benchmark later this year (announced at NT symposium).
    
    >>>
    When memory becomes sharable between processors we usually think of SMP
    as being the solution, not clusters..
    >>>
    
    See prev paragraph on VLM, but you raise an interesting point ie. in
    a Wolf Pack cluster where each node is handling a specific set of
    resources, a SMP node may not be a good solution. Reason ? In a SMP
    environment, if each cpu is accessing/updating a common area of
    memory, then there is going to be lot of cache invalidation activity
    between cpu's. This means the local cpu's will be going to main memory
    and/or disk much more often.
    
    We shall see ..
    
    Regards,
    
    / Kerry

	Just read the latest TechNet from MS re: SQL in a clustered environment.

	It isn't clear, but when they start talking about SQL becoming highly
	parallelized (ugly word, isn't it?) it seems to imply that they will
	be accessing a single SQL database from multiple nodes in a cluster.
	Maybe I misread, and of course it is way out in the future, but if you
	haven't read the latest TechNet, it might shed some other light on
	the situation.  I tend to agree with Kerry; if MS expects to compete
	with Oracle on its ability to scale, MS will have to some up with some
	kind of DLM.  I am really interested to see how MS will rise to the
	competition that Oracle will provide when NT5.0 comes out on Alpha.

tgc

    >>> The point is that VMS DB applications like RDB and Oracle have been 
    taught to use DLM to coordinate access to shared data.  There aren't any 
    such NT applications (yet)... 

Actually, VMS applications simply know how to cope with multiple processes
accessing the same data.  They use the lock manager to cooperate.  When
you run them on a cluster, the lock manager becomes distributed, but the
applications don't need to know anything about it.

How do separate processes cooperate on a single node on NT?  Why can't that
scale to a cluster the same way it does in VMS?

>Actually, VMS applications simply know how to cope with multiple processes
>accessing the same data.  They use the lock manager to cooperate.  When
>you run them on a cluster, the lock manager becomes distributed, but the
>applications don't need to know anything about it.

    In my experience a large majority of applications actually don't know
    or care about the locking when accessing data, but just use RMS which
    does the locking for them - effectively the applications punts
    responsibility to the filesystem/programming API.
    
> How do separate processes cooperate on a single node on NT?  
    
    That I cannot say.  NT filesystems do export primitives to allow this
    to take place (range locks and oplocks), but I do not know whether they
    are used at all or whether people rely on general synchronisation
    primitives like semaphores and whatnot..
    
>    Why can't that scale to a cluster the same way it does in VMS?
    
    In a word, because of write behind caching (actually any caching). 
    When OpenVMS clusters came out the world was used to write through and
    no caching and so everyone knew that every $QIO was going to the oxide and
    codes and planned accordingly.  Hence the delay in a QIO being serviced
    was predictable and large.
    
    These days every operating system relies on caching to provide
    performance. Application writers may only be vestigially aware of
    caching (one hopes that they understand the crash-consistency
    contraints).  As soon as write behind caches hit OpenVMS clusters the
    time to service a IO changes from being predictable and slow to being
    unpredictable and anything between lightning fast (returned from
    memory) and excrutiatingly slow (remote flush and read from the oxide).
    Trust me, I was there during Spiralog development.
    
    The unpredictablbility depends on whether the last write was on this
    node or another one.  So adding an n'th application to second node will
    not be anything like the scaling advantage that adding it to the same
    node (in which case all writes are to cache). 
    
    This is not my argument, this is Tom Miller's (NT FS architect) and 
    Rod Gamache's (NT cluster architect).  Yes there are lots of arguments
    against this but you have to go down to the next level of detail and to
    some degree assume that one has control over WIN32.
    
    [To restate my position, I believe in shared everything, but I
     would not like to see us careening down a path which is  "obvious to
     all of us", unprepared to argue through a well thought out and
     differing view.  The needs of `dynamic load balancing' as enunciated
     earlier is (I hope) unasailable, but the `it's obviously scalable' is
     a harder sell), as is the as yet unmentioned `more obvious
     availability paradigm'
    
     As it turns out tomorrow is my last day here so it's not important to 
     me, but for DEC's (sic) sake I urge that this high quality conversation 
     continues 
    
     rod]