AMD Processors
Decrease font size
Increase font size
Topic Title: more memory!?
Topic Summary:
Created On: 03/10/2005 07:36 AM
Status: Read Only
Linear : Threading : Single : Branch
1 2 Next Last unread
Search Topic Search Topic
Topic Tools Topic Tools
View similar topics View similar topics
View topic in raw text format. Print this topic.
 03/10/2005 07:36 AM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

Is anyone out there making AMD64 systems with external/auxiliary memory controllers? "Scaling memory with the number of CPUs" doesn't make sense for a data-driven class of applications. E.g., we have a customer looking for a system that can cache an entire terabyte database in RAM. The database query rate can easily be handled by a few CPUs, it's the I/O rate that's the bottleneck. Caching the whole thing in system RAM would eliminate the I/O bottleneck.

Even supercomputers like the Cray XD1 take the same approach - to increase system memory, you have to add processor boards. It appears that the only way to remedy this bottleneck with an Opteron-based system is to go with an external RAMdisk (like Texas Memory Systems or Imperial Technology), but that still means going through an external I/O bus and imposing a fair amout of latency. (Oh yes, we need sub-millisecond responses to the queries, obviously. Our current design can respond in 500 microseconds when the data is all in RAM. But as the data set grows, paging in and out of an external storage system is going to be a problem.) To get a Cray system with 1TB of RAM we would have to have 144 processors installed. Rather silly since only 2 processors would be needed to handle the expected query rate.

The SGI Altix claims to support independent scaling of memory vs CPUs and that may wind up being our target platform, much as I think the Itanium is a mistake.

This is certainly an extreme case, but I think it's pretty obvious that system designers need to wake up to the fact that CPU speed is no longer a major problem in solving large problems - having access to the data is. So-called "balanced" designs are missing the point; I have 2GB of RAM on my laptop already. There's a huge class of problems I can easily solve on this laptop now. The only reason to even consider a supercomputer is when the class of problem is orders of magnitude larger than I can easily handle on a laptop or desktop machine. And yet, the amount of memory available to a single processor in a so-called supercomputer is *not* orders of magnitude greater than my laptop. 8GB vs 2GB is nothing, and 8GB is nothing for a 64 bit machine.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 03/10/2005 10:43 AM
User is offline View Users Profile Print this message

Author Icon
Calvn_Swing
Senior Member

Posts: 293
Joined: 01/13/2004

Well, I don't understand why you think that the limit is 8 Gigs. Most dual opteron systems have 8 dimms. They make 2 and 4 gig sticks of ram - expensive, yes - so that means up to 32 Gigs of ram.

That being said, it is still no where near your desired capacity of 1 Terabyte. So it is still a good point. However, I would say that the balanced approach is only questionable in certain areas of application - yours for example. Because in DCC, it is still the processor that is the limitation. I don't fill up my 1Gig of RAM with the small scenes I work on in Maya. I am planning on putting in 4 Gigs when I upgrade so I can work on city scale scenes a but more effectively. That will be more than enough. However, the two Opteron CPU's can take several hours - three weeks for my last 30 second animation - to process a single frame. Point being, the balanced approach is a way to balance the demand of customers like you and I for performance gains. You need more memory, I need more speed. So they do both.

None the less, loving opterons as much as I do, I hope you find a solution somehow. Surely there is a motherboard manufacturer that has a board for opterons that has more than 8 dimm slots!!! However, take your Terabyte of RAM and assume that you are buying 2 Gig sticks. You are going to need 500 sticks of RAM? That is a lot of power and more importantly a lot of space! AND, as the size of the RAM increases the latency also increases doesn't it? And Opterons have interated memory controllers. I don't imagine that with 500 sticks of 2 Gig Ram that you'd have much of a performance increase over a RAID array of SCISI drives, but it would be a lot more expensive!!!

But anyway, it's fun to think about a Terabyte of RAM.

adios...

Just my thoughts on it...

-------------------------
Iwill Zmax DP SFF Barebones kit... (300W PSU, 8 USB, 1 FW, etc...)
Dual AMD Opteron 275's | 2 Gigs (2 Dimms 1024Mb) Patriot RAM 3-3-3-8
WD Caviar 80 Gig HDD - System
WD Raptor 74 Gig HDD
ATI Radeon 9800 Pro 256
NEC ND-3520A Dual Layer DVD±RW | Mitsumi Floppy/Card Reader combo
 03/10/2005 11:33 AM
User is offline View Users Profile Print this message

Author Icon
redclawkefar
Senior Member

Posts: 1082
Joined: 02/15/2004

QUOTE (Calvn_Swing @ Mar 10 2005, 10:43 AM) Well, I don't understand why you think that the limit is 8 Gigs. Most dual opteron systems have 8 dimms. They make 2 and 4 gig sticks of ram - expensive, yes - so that means up to 32 Gigs of ram.

That being said, it is still no where near your desired capacity of 1 Terabyte. So it is still a good point. However, I would say that the balanced approach is only questionable in certain areas of application - yours for example. Because in DCC, it is still the processor that is the limitation. I don't fill up my 1Gig of RAM with the small scenes I work on in Maya. I am planning on putting in 4 Gigs when I upgrade so I can work on city scale scenes a but more effectively. That will be more than enough. However, the two Opteron CPU's can take several hours - three weeks for my last 30 second animation - to process a single frame. Point being, the balanced approach is a way to balance the demand of customers like you and I for performance gains. You need more memory, I need more speed. So they do both.

None the less, loving opterons as much as I do, I hope you find a solution somehow. Surely there is a motherboard manufacturer that has a board for opterons that has more than 8 dimm slots!!! However, take your Terabyte of RAM and assume that you are buying 2 Gig sticks. You are going to need 500 sticks of RAM? That is a lot of power and more importantly a lot of space! AND, as the size of the RAM increases the latency also increases doesn't it? And Opterons have interated memory controllers. I don't imagine that with 500 sticks of 2 Gig Ram that you'd have much of a performance increase over a RAID array of SCISI drives, but it would be a lot more expensive!!!

But anyway, it's fun to think about a Terabyte of RAM.

adios...

Just my thoughts on it...
some things some people can really hit the nail on the head... WELL SAID calvin.

i'd have to agree with you, no way around it.... unless of course someone gave me a free 1 terabyte of ram and all the hardware to go with it... and ya, then i just wouldn't care what people said... it'd be all bragging rights.

-------------------------
SYSTEM-1: 2x Viewsonic VX900-2 19"| Tyan S2885ANRF (Rev.A, bios v2.07) | 1x ATI Radeon X800 Pro 256MB | 2x AMD Opteron 248-HE Single-Core CPU | 8x 512mb PC3200 Corsair XMS Series ECC Reg. (CMX512RE-3200LL) | 2x Seagate LVD 36.7Gb U320 SCSI 10,000rpm RAID0 (ST336607LW) | 1x Seagate LVD 36.7Gb U320 SCSI 1
 03/10/2005 12:37 PM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

Thanks for the response, Calvin.

Yes, I understand that we're two opposite ends of the curve. The point is that there appears to be only one company out there (SGI) selling hardware that allows you to scale processors and memory independently. And yes, I'm an AMD64 fan too, I really would be rather disappointed to have to go with an Itanium solution.

And yes, I understand that Opterons have integrated memory controllers. My very first question in the post was whether anybody makes external/auxiliary memory controllers though, because the Opteron controller is (a) limited to 8 DIMMs and ( slows down as memory size goes up.

It's pretty clear that the memory limitations here are due to the on-chip implementation, you just can't put big beefy line drivers onto the chip along with everything else competing for space there. The Opteron controller only has so many watts to spend on the memory bus. As the number of addressable DIMMs goes up, the effective voltage available per byte goes down, and thus you have to run slower and slower memory to keep operating.

But with an off-chip / auxiliary memory controller that had its own power supply, there would be no reason for memory size to degrade memory speed. The ideal would be a module with >= two HyperTransport ports that was nothing but a dedicated memory controller. At 6.4GB/sec you would still have as much bandwidth as the direct-attached RAM, just with higher latency.

re: 500 sticks of RAM - well, yeah, that's a separate problem. 4500Watts of power right there. It would probably be best partitioned up into individual modules, maybe 32 DIMMs / 128GB per module. But the point is to even have the option of installing a memory-only module into a big Opteron server rack. Right now that option doesn't even exist.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 03/13/2005 07:54 PM
User is offline View Users Profile Print this message

Author Icon
OrangesAway
Senior Member

Posts: 202
Joined: 03/09/2004

At the moment, no one makes a memory only or memory specific expansion for Opteron systems. It is certainly possible, at least inter-board (maybe not inter-chassis) with something like the IWILL approach for connecting two CPU boards.

The AMD 64 architecture, in 64 bit mode, supports a full 64-bit virtual address space with 4 petabytes of physical memory (2^52 or 52 bits) -- However the initial implementations (Opteron, Athlon FX) support 256 TB (2^48 or 48-bit) virtual memory addresses and 1 TB (2^40 or 40-bit) physical memory addresses. So even if a system had memory expansion, it would "just" support your requirement.

I would bet by the time you build an Altix system with 16 M-bricks at 64 GB each, plus the R-brick, C-brick and other equipment, you will find it still much cheaper to go with one of the solid state disk systems. In any case, if your requirements are that strict, then that may be your only choice.

Texas Memory' ">http://www.superssd.com/products/tera-ramsan/indexb.htm qoutes "Latency: Less than 20 microseconds" and Imperial Technology' ">http://www.imperialtechnology....roducts_megaram10k.htm qoutes 35 microseconds. Naturally have to count some additional overhead, but would assume your 500 microsecond target is reachable.

Otherwise, in general agree with you on the need to scale the memory. I think the industry agrees with the problem and thats why you see proposals like FB-DIMM, see MemForum FB DIMM' ">http://www.memforum.org/webcasts/fbd/fbd.

 03/13/2005 08:50 PM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

Thanks for the info, OrangesAway. That FB-DIMM info looks really promising, it'll be interesting to see what comes of it.

And yeah, I'm beginning to think the external RAMdisk approach is the more practical route for now.


-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 03/15/2005 01:53 PM
User is offline View Users Profile Print this message

Author Icon
RedEXE
Senior Member

Posts: 213
Joined: 10/25/2004

QUOTE (Calvn_Swing @ Mar 10 2005, 07:43 AM) Well, I don't understand why you think that the limit is 8 Gigs. Most dual opteron systems have 8 dimms. They make 2 and 4 gig sticks of ram - expensive, yes - so that means up to 32 Gigs of ram.

That being said, it is still no where near your desired capacity of 1 Terabyte. So it is still a good point. However, I would say that the balanced approach is only questionable in certain areas of application - yours for example. Because in DCC, it is still the processor that is the limitation. I don't fill up my 1Gig of RAM with the small scenes I work on in Maya. I am planning on putting in 4 Gigs when I upgrade so I can work on city scale scenes a but more effectively. That will be more than enough. However, the two Opteron CPU's can take several hours - three weeks for my last 30 second animation - to process a single frame. Point being, the balanced approach is a way to balance the demand of customers like you and I for performance gains. You need more memory, I need more speed. So they do both.

None the less, loving opterons as much as I do, I hope you find a solution somehow. Surely there is a motherboard manufacturer that has a board for opterons that has more than 8 dimm slots!!! However, take your Terabyte of RAM and assume that you are buying 2 Gig sticks. You are going to need 500 sticks of RAM? That is a lot of power and more importantly a lot of space! AND, as the size of the RAM increases the latency also increases doesn't it? And Opterons have interated memory controllers. I don't imagine that with 500 sticks of 2 Gig Ram that you'd have much of a performance increase over a RAID array of SCISI drives, but it would be a lot more expensive!!!

But anyway, it's fun to think about a Terabyte of RAM.

adios...

Just my thoughts on it...
It took you a week to render One 30 sec's Of animation in maya? WOW nothing has ever taken my dual 250's that long!!! What was the face cout.. Id love to see a image of that scene...
 03/21/2005 10:21 PM
User is offline View Users Profile Print this message

Author Icon
functional-pc
Senior Member

Posts: 1558
Joined: 01/10/2005

http://www.bitmicro.com/products_edisk_35_scsin.php' ">http://www.bitmicro.com/products_edisk_35_scsin.php

good enough for the military...

-------------------------
Latest workstation:
Lian li PC-A05B, KN3-SLi, 4400 Brisbane, Freezer 64 Pro, 2 gigs Super-Talent T800UX2GC4, 36 raptor system drive, Caviar RE 320 storage drive, Hiper 580, Ati 1300 Pro, running Vista Home Premium 64-bit with Symantec 10.2.224
 03/21/2005 10:38 PM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

Flash disks are intended for a pretty different purpose. Those things peak at 20MB/sec, which is much slower than a decent desktop hard drive. The external DRAM units are many times faster than rotating hard drives. Flash disks are intended for installations where ruggedness is more important than speed, and that's not the casefor my situation.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 03/22/2005 12:26 AM
User is offline View Users Profile Print this message

Author Icon
functional-pc
Senior Member

Posts: 1558
Joined: 01/10/2005

i just figured the .02 msec access time would have to help you somehow...sorry

-------------------------
Latest workstation:
Lian li PC-A05B, KN3-SLi, 4400 Brisbane, Freezer 64 Pro, 2 gigs Super-Talent T800UX2GC4, 36 raptor system drive, Caviar RE 320 storage drive, Hiper 580, Ati 1300 Pro, running Vista Home Premium 64-bit with Symantec 10.2.224
 03/22/2005 03:44 AM
User is offline View Users Profile Print this message

Author Icon
GNUS inc
Member

Posts: 136
Joined: 10/07/2003

highlandsun
If you are looking for such amounts of RAM, you are probably looking for appropriate amount of CPU power - I can't imagine task that require 1TB or RAM and only two CPU.
RAM is limited by physical possibilities of DDR devices per channel - only 4 sticks of DDR400 per dual-channel controller or 8 DDR266.
And integrated controller in CPU allow to scale RAM only with CPU.

If you want Opterons - look for a cluster (on InfiniBand or Cray), only 8-way machines are available and bigger systems are not expected this year.

Solid state systems are not like memory, it's like disk storage - they are not the solution you are looking for.

-------------------------
AMD rulezzzzz!
www.etegro.com
 03/23/2005 10:31 AM
User is offline View Users Profile Print this message

Author Icon
jason57570
Senior Member

Posts: 707
Joined: 10/07/2003

lord of the rings was rendered on A64 quad based opteron systems, like 500 of them. I heard they ran red hat linux as it was 10x's as fast as windows..

they first where using intel, but switched to amd servers/workstations after part 1 was it

they could even render in realy time some stuff

-------------------------
Ati x1900 Crossfire & FX-60 Dual Core
[3DMark 2006=12,573
[3DMark 2005=20,667
[3DMark 2003=42,891
 04/06/2005 02:36 AM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

I understand that a solid state disk is not the same as main memory. But for a database application like this, it might be good enough. It would essentially be a large filesystem cache, to reduce some of the latency caused when the main memory database cache is filled and has to page things out.

Modifying the original question a little bit - the requirement is to be able to scale up to a 1TB size within 2-3 years. So the system we need to have today can be much smaller, as long as there is a clear upgrade path to the required size down the road.

I understand that the Opteron memory controller is integrated with the processor, and that is why typical system designs scale memory with the number of CPUs. But there's nothing inherent to the Opteron that says it can only be done this way. Like I suggested above, there's all this Hypertransport bandwidth available on multiple channels, you could easily design a system that includes a Hypertransport node that consisted solely of a dedicated memory controller, and have it provide access to all the other HT-attached processors. It would be higher latency than directly attached memory, but it would still be faster than anything external, and fundamentally no different than systems today where multiple processors access each other's memory across HT.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 04/27/2005 04:14 PM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

Looks like Samsung already developed an 8GB DDR DIMM
http://www.samsung.com/Product...nduc..._0000059139.htm' ">http://www.samsung.com/Product...0040624_0000059139.htm

I wonder when it will reach the market, if at all. I suppose DDR2 or DDR3 are getting all the attention now. They don't even list 1Gbit FBGA components on their web site, the 1Gbit parts they list are only TSOP.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 04/27/2005 10:31 PM
User is offline View Users Profile Print this message

Author Icon
Calvn_Swing
Senior Member

Posts: 293
Joined: 01/13/2004

Well, at 8Gigs you're a little closer!

Still looking for the system though? That at least means no itaniums yet I guess...

I ran across a SUN Opteron workstation the other day with a connection for a daughtercard that hosted a second processor. I know it isn't a memory controller, but theoretically you could create a memory controller and hook it up to the same connector couldn't you? I wanted to ask when I first saw this post if the volume of the project was big enough to justify custom boards...

Might be worth looking into. Might not (only one processor if the daughtercard doesn't have the second).

best of luck!

-------------------------
Iwill Zmax DP SFF Barebones kit... (300W PSU, 8 USB, 1 FW, etc...)
Dual AMD Opteron 275's | 2 Gigs (2 Dimms 1024Mb) Patriot RAM 3-3-3-8
WD Caviar 80 Gig HDD - System
WD Raptor 74 Gig HDD
ATI Radeon 9800 Pro 256
NEC ND-3520A Dual Layer DVD±RW | Mitsumi Floppy/Card Reader combo
 04/28/2005 12:55 AM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

I'm trying to cover as many options as possible. We've already done some tests on one Itanium-based platform, we are looking at a couple other Itanium systems still. I don't want to make a conclusion/recommendation based on only one or two data points. At present it still looks like there are no Opteron based servers that fit the bill, tho as you say, 8GB DIMMs would bring it a lot closer.

I'm pretty sure our client wants an off-the-shelf solution. They'll only be deploying a small number of these things, I dunno if we could justify the price of custom development. The question of support would be rather thorny then, both software and hardware maintenance.


-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 04/28/2005 12:48 PM
User is offline View Users Profile Print this message

Author Icon
jes
Senior Member

Posts: 1134
Joined: 10/22/2003

Just out of interest, what sort of database is it? I'm assuming since you only need it cached in ram that it's a DW than a transactional database.

-------------------------
The opinions expressed above do not represent those of Advanced Micro Devices or any of their affiliates.
http://www.shellprompt.net
Unix & Oracle Web Hosting Provider powered by AMD Opterons
 04/28/2005 05:55 PM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

It's actually an LDAP directory, (OpenLDAP 2.3) with a transactional backend. Writes will be synchronous, and are expected to be relatively slow, but reads must be fast.

To be quite honest, this is not a processor-intensive job by any stretch of the imagination. My development machine, an AMD64 3000+ Winchester can process around 2000 operations per second (mix of reads and writes) when all the data is in memory so we have absolutely no worries about the processor load; we know that any decently fast machine will be able to meet our target in that respect. But response latency is what we're trying to eliminate, and the answer to that is As Much Cache As Possible.


-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
 04/29/2005 03:30 AM
User is offline View Users Profile Print this message

Author Icon
jes
Senior Member

Posts: 1134
Joined: 10/22/2003

You have an LDAP directory with a 1TB footprint? In my opinion, you'd be better splitting the workload horizontally rather than vertically, i.e. more lower capacity servers rather than one (I *really* hope you don't just have one) "superserver".

QUOTE
t's actually an LDAP directory, (OpenLDAP 2.3) with a transactional backend. Writes will be synchronous, and are expected to be relatively slow, but reads must be fast.


I'm not quite sure I follow this bit, if the writes are written to disk (and with a 1TB footprint they *really* should be written to disk), then how can you be sure that reads are consistent?

True story - I worked at a site where they used a *really* large Oracle database, it was so large that people used to complain that their queries ran too slow. Rather than looking at the problem "properly", they (the management) decided to switch to a (name withheld to protect the guilty) competitors product which held the database completely "in-memory" (to me database and "in-memory" are at odds with each other). To cut a long story short, the users were completely in awe of how fast the new database was, their query times were *WAY* down......everyone was happy....right till the day when the box died, the "cost" to the busines of that lost data (i.e. data that had never been commited to disk) was incalculable.....but it far outweighed any query time savings they'd made.

They went back to Oracle rapidly....

-------------------------
The opinions expressed above do not represent those of Advanced Micro Devices or any of their affiliates.
http://www.shellprompt.net
Unix & Oracle Web Hosting Provider powered by AMD Opterons
 04/29/2005 05:47 AM
User is offline View Users Profile Print this message

Author Icon
highlandsun
B1FF

Posts: 111
Joined: 03/10/2005

QUOTE You have an LDAP directory with a 1TB footprint? In my opinion, you'd be better splitting the workload horizontally rather than vertically, i.e. more lower capacity servers rather than one (I *really* hope you don't just have one) "superserver".

That makes little sense. Just as with RAID, when you use a set of machines instead of a single machine, your MTBF for the total system is inversely proportional to the number of machines. Not to mention the increased administration overhead from managing the larger number of servers, and the painstaking effort to apportion the data to each server and keep the raw data volumes (as well as query volume) balanced as the data grows over time.

There will be 10-20 of these servers located at geographically distributed sites. Certainly not just one. The sysadmin overhead will be bad enough at that. If each site is a farm of smaller servers, the sysadmin cost will skyrocket.

QUOTE
I'm not quite sure I follow this bit, if the writes are written to disk (and with a 1TB footprint they *really* should be written to disk), then how can you be sure that reads are consistent?


Excuse me? How does a filesystem buffer cache allow you to have consistent reads when it operates in writethru mode? This is such a basic function of any cache, that's not even a worthwhile question.

There are four basic guarantees in a transactional database - Atomicity, Consistency, Isolation, and Durability. Atomicity means that no matter how many updates are performed in a single transaction, either all of them succeed or all of them fail when the transaction is committed. Consistency means that a transaction creates a new and valid state of data, or if any failure occurs, returns all data to its state before the transaction started. Isolation means that a transaction in progress is isolated from any other transaction - partial updates in one transaction are not visible in any other transaction. Durability means that committed data is saved by the system such that, even in the event of a failure and system restart, the data is available in its correct state.

Your question has mainly to do with Durability. In this case the standard approach used is write-ahead logging - all database updates are first written to a log file as they occur, then written (through the cache) to the actual database files. When the commit happens, each database file is stamped with the last logfile position corresponding to its current state.

If a system failure occurs before the commit completes, a consistent database state is easily restored at the next startup. An automatic recovery procedure reads the logs and database state. Any transactions that were in progress but not committed in the logs are rolled back, and any transactions committed in the log but not fully written to the database are completed.

All of our database writes are transactional; while caching is used extensively for both reads and writes the database guarantees that all writes have been flushed to the operating system and that the operating system buffers have been sync'd before the Commit function returns to the application.

So as I said originally, our writes are fully synchronous and we expect them to be slow, and we accept that.

For the general question "how do you know your reads are consistent" - fine grained locking is used everywhere. Items that are currently being modified will cause read operations accessing those items to block, but that's life. In general, the usage profile for an LDAP directory is many reads and few writes; most read requests will be able to be satisfied immediately without blocking behind any writes.

-------------------------
Asus A8V Deluxe, Opteron 185, 4xSamsung 1GB Reg/ECC DDR400, AC Freezer64 Pro, Antec Sonata case w/TruePower480
Statistics
112018 users are registered to the AMD Processors forum.
There are currently 0 users logged in.

FuseTalk Hosting Executive Plan v3.2 - © 1999-2014 FuseTalk Inc. All rights reserved.



Contact AMD Terms and Conditions ©2007 Advanced Micro Devices, Inc. Privacy Trademark information