Sign in with
Sign up | Sign in
Your question

AnandTech Database benchmark

Tags:
  • CPUs
  • Benchmark
  • Database
Last response: in CPUs
Share
April 4, 2002 6:23:54 AM

<A HREF="http://www.anandtech.com/cpu/showdoc.html?i=1606&p=1" target="_new">http://www.anandtech.com/cpu/showdoc.html?i=1606&p=1&lt;/A>

Nice examination of HyperThreading in relation to database software. Rather odd that it actually hurt performance in one test...

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?

More about : anandtech database benchmark

April 4, 2002 2:42:32 PM

Thanks for the link Kelledin. I dont believe I have ever seen duals go head-to-head in a "real world" environment. Im curious why Itanium wasnt used. I thought Itanium was for high end servers and what not. Or am I on crack about what Itanium is for?
Anywho, most folks I have talked to agree that Xenon without HT is crap. With HT enabled, it competes very well. I really like Anand's explanantions as well. As opposed to some other reviewers *cough* Tom *cough*.


Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.
April 4, 2002 3:40:22 PM

Itaniums are $3000 each, just for the processor. I doubt Anand has the cash for that, and I doubt Intel is frivilously throwing them at reviewers.

<font color=blue>If you don't buy Windows, then the terrorists have already won!</font color=blue> - Microsoft
Related resources
April 4, 2002 6:09:01 PM

Quote:
Thanks for the link Kelledin. I dont believe I have ever seen duals go head-to-head in a "real world" environment. Im curious why Itanium wasnt used. I thought Itanium was for high end servers and what not. Or am I on crack about what Itanium is for?

The Itanium is targetted for the scientific community where floating point performance is most important. It is not designed to be in a web server, file server, or desktop machine.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 4, 2002 6:30:46 PM

Nice benchmark i hope anandtech can make more bench of duallie.Under differente OS chipset and others
cheap, cheap. Think cheap, and you'll always be cheap.AMD version of semi conducteur industrie <P ID="edit"><FONT SIZE=-1><EM>Edited by juin on 04/04/02 02:36 PM.</EM></FONT></P>
April 4, 2002 6:48:13 PM

I thought it was a good article and unbiased. Excellent review.
April 4, 2002 7:19:40 PM

Anandtech is the <i>creme de la creme</i> of hardware reviews, hey I can't fault the guy ! Concise, accurate and unbiased reviews.

As for the artical I didnt realise how powerful a Dual Athlon system was until now !...makes me wonder how fast Sledgehammer will be.

Intels hyperthreading also looks very good ! Do applications need to be optimised for Hyperthreading to benifit from it ? cause Anandtech didn't say anything about it. Do servers running standard applications automatically benifit from Hyperthreading ?

<font color=purple>~* K6-2 @ 333MHz *~
I don't need a 'Gigahertz' chip to surf the web just yet ;-)</font color=purple>
April 4, 2002 7:22:52 PM

Interesting, but I do have one observation here. They used a dual-channel DDR SDRAM chipset for the Xeons. This uses PC1600 DDR SDRAM, which has hideous latency compared to just about anything else. The AMD-based system used PC2100 DDR SDRAM, which would have considerably less latency. Adding more memory channels does not decrease latency.

I would have preferred a review that used the i860 chipset, which uses dual-channel RDRAM like the i850 and supports up to two Xeon processors.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 4, 2002 7:51:59 PM

I thought Dual Channel decreases latency, which is why RDRAM needed that besides bandwidth?

--
For the first time, Hookers are hooked on Phonics!!
April 4, 2002 7:54:12 PM

Dual-channels increase the available bandwidth. If you want to fetch a byte it will still travel at PC1600 speeds from memory to processor. Multiple channels will not help that. It can only travel across a single channel.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 4, 2002 8:00:04 PM

Then I suggest you e-mail the editor and tell him you'd like a different comparison. Strange though, we didn't get to see normal benches of single Xeon, no SMP enabled, using normal tasks compared to a normal P4 2.2GHZ, to see how Dual Channel DDR does better or not than RDRAM currently.

--
For the first time, Hookers are hooked on Phonics!!
April 4, 2002 8:04:09 PM

I learned long ago that emailing editors was a waste of time. They get so many emails that they rarely spend more than a few seconds glancing over yours. Even if they do read it, they are not going to redo a whole article simply because I want them to do it. I have emailed some of them, including editors here at THG, and have received absolutely no response.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 4, 2002 10:42:59 PM

Is it just me, or is that orange and blue P4-M ad banner absolutly ugly? I mean.... ugh. Not that the Radeon ones were any better. *Sigh*

Anywho, in the article, Anand mentioned that he would do more testing later with a whole article dedicated to the E7500. Granted, the i860 chipset might have been better, but I think that Anand's point is that for the market point that he's looking at (entry-level to mid-level server market), and since I have no clue as too the pricing, I'd have to assume there is a significant difference between the i860 and E7500 in price and target markets. Also, if I'm to understand Intel's roadmap correctly (as gleaned from <A HREF="http://freespace.virgin.net/m.warner" target="_new">http://freespace.virgin.net/m.warner&lt;/A>), DDR memory is going to be the route taken for Xeon processors. Also, they used DDR266 sticks:
Quote:
The test beds were configured with 2GB of DDR266 SDRAM using two 1GB sticks. Since the E7500 chipset requires memory to be installed in pairs we had to use multiple DIMMs. Also remember that the dual-channel memory controller of the E7500 chipset only ran the memory bus at DDR200 speeds.

So, the limiting factor would have been the chipset, not the memory. And finally, there is still 3.2GB/s of bandwidth with either the i860 or E7500 (according to Anandtech). All in all, I think it was an unbiased article, biased only by the equipment availible.

-SammyBoy
April 5, 2002 12:09:29 AM

Ray now it's been the second time THG reports that Intel has officially said that by the end of the year, Dual Channel DDR is the choice for the P4 later on. Maybe you'll say differently but I'm starting to heavily beleive that, as it has been talked widely in the web. What can you say about it? It seems indeed that Rambus will stay for a little but not as much as we expected with Intel.

--
For the first time, Hookers are hooked on Phonics!!
April 5, 2002 1:21:55 AM

Quote:
Interesting, but I do have one observation here. They used a dual-channel DDR SDRAM chipset for the Xeons. This uses PC1600 DDR SDRAM, which has hideous latency compared to just about anything else. The AMD-based system used PC2100 DDR SDRAM, which would have considerably less latency. Adding more memory channels does not decrease latency.

1) As you pointed out several months ago, 1GB and greater RIMMs are hard to come by. This is a critical shortcoming when it comes to database servers.

2) Dual-channel or not, everything we've heard from everyone but Intel suggests that PC1600 would have <i>lower</i> latency than PC800.

3) Didn't you say a while back that DDR's latency increases as you increase the bandwidth?

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?<P ID="edit"><FONT SIZE=-1><EM>Edited by kelledin on 04/04/02 08:24 PM.</EM></FONT></P>
April 5, 2002 1:22:07 AM

They may have used PC2100 DDR SDRAM sticks, but they were running at 100MHz, as PC1600. That is how the chipset works.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 5, 2002 1:54:17 AM

Quote:
Ray now it's been the second time THG reports that Intel has officially said that by the end of the year, Dual Channel DDR is the choice for the P4 later on. Maybe you'll say differently but I'm starting to heavily beleive that, as it has been talked widely in the web. What can you say about it? It seems indeed that Rambus will stay for a little but not as much as we expected with Intel.

Intel is currently agnostic about memory technology. A dual-channel PC2100 DDR SDRAM chipset is coming by the end of the year, but it is only playing catchup to the RDRAM chipset. Dual-channel PC1066 RDRAM with a 533MHz FSB has less latency than dual-channel PC2100 DDR SDRAM on a 533MHz FSB. Unless you need more than 2GB of memory, I recommend the better-performing RDRAM chipset.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 5, 2002 2:03:40 AM

Quote:
1) As you pointed out several months ago, 1GB and greater RIMMs are hard to come by. This is a critical shortcoming when it comes to database servers.

This is true. We need some motherboards that implement more than 2 RIMM slots per channel, or 1GB RDRAM modules.

However, this benchmark only used 2GB of memory! An i860-based RDRAM platform would have worked just fine.


Quote:
2) Dual-channel or not, everything we've heard from everyone but Intel suggests that PC1600 would have lower latency than PC800.

Can you provide a link?


Quote:
3) Didn't you say a while back that DDR's latency increases as you increase the bandwidth?

DDR's latency per clock does increase as you increase its bandwidth. It's real-time latency decreases as each clock takes less time to complete. RDRAM's latency per clock remains the same as you increase its bandwidth, ignoring other factors that might contribute per-clock latency, such as asynchronous memory-to-FSB speeds.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 5, 2002 2:22:15 AM

Is there a reasonable excuse to actually purchase an Itanium at 3k a pop? I could buy two decent AXP systems for that much!

"When there's a will, there's a way."
April 5, 2002 2:35:07 AM

Quote:
Can you provide a link?

DDR has lower latency than RDRAM; that's rather common knowledge by now (unless you can provide a refuting link?) There hasn't been much in the way of investigating PC1600, however, so no, I have no link.

Also, dual-channel memory apparently <A HREF="http://www6.tomshardware.com/mainboard/01q2/0106041/nfo..." target="_new">does decrease latency</A>, with either RDRAM or DDR.

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?
April 5, 2002 2:46:32 AM

PC2100 DDR SDRAM has lower latency than PC800 RDRAM. PC1600 does not. As far as the nForce, it decreased latency by offering more bandwidth. This reduced congestion. There is no difference in bandwidth offered between the i860 and the E7500 chipsets though.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 5, 2002 2:48:20 AM

If you need a computer system to perform scientific floating-point calculations, and you are working with massive amounts of data, then you need a 64-bit processor with the best floating point performance. Itanium offers the best price/performance ratio of 64-bit processors.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
April 5, 2002 3:00:30 AM

Quote:
There is no difference in bandwidth offered between the i860 and the E7500 chipsets though.

There <i>is</i> a difference between single-channel and dual-channel PC1600 though, and it's more than just bandwidth. Being able to have twice as many pages open (via dual channels) nicely cuts down on latency by reducing the need for the memory controller to repeatedly open and close various pages of memory.

As for the i860 chipset...

Sure, AnandTech's test didn't require more than 2GB RAM, but there's a great many database situations where more than 2GB is required. It would be a disservice to review a platform that much of the target audience couldn't even use.

Regarding the more-slots-per-channel, why doesn't Intel make a suitable i860 reference board with more RIMM slots? Third party manufacturers seem reluctant. Plus, I think we all remember why the i820 had to have its third RIMM slot dismembered.

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?
April 5, 2002 3:21:35 AM

The new XEON MP chips top that - a 1.6GHz XEON MP will run you over US$3600.

Both the intel (i7500) and ServerWorks (Grand Champion) XEON MP solutions use dual channel DDR.

The Serverworks chipset lets you set up your memory much like RAID solutions. You can stripe it for maximum performance, mirror it (and hot swap failed DIMMS), or just use it as one big block.

- JW
April 5, 2002 1:54:39 PM

In case you haven't found an answer for your question regarding Hyperthreading, programs do have to be optimized for it. They did mention in the Anandtech article that Hyperthreading requires the optimizations. One of their benchmarks even showed ~5% loss of performance.
April 5, 2002 4:42:41 PM

Quote:
Is there a reasonable excuse to actually purchase an Itanium at 3k a pop? I could buy two decent AXP systems for that much!


Itaniums are designed for large scale machines, not the home desktop. There is no reason for you to buy an Itanium, period. Even if they were $5 each I wouldn't buy one (well, I would, but just for a keychain).

<font color=blue>If you don't buy Windows, then the terrorists have already won!</font color=blue> - Microsoft
April 5, 2002 4:43:35 PM

Quote:
DDR has lower latency than RDRAM; that's rather common knowledge by now (unless you can provide a refuting link?) There hasn't been much in the way of investigating PC1600, however, so no, I have no link.


Look at the memory section, apparently RDRAM has lower latency than DDR, if the results are to be believed.

<font color=blue>If you don't buy Windows, then the terrorists have already won!</font color=blue> - Microsoft
April 5, 2002 7:36:25 PM

Quote:
The AMD-based system used PC2100 DDR SDRAM, which would have considerably less latency. Adding more memory channels does not decrease latency.


The hell it dosent, the nforce has lower latency than the kt266a, the dual channel i850 has lower latency than the rdram p3 chipset.

:wink: The Cash Left In My Pocket,The BEST Benchmark :wink:
April 5, 2002 7:49:56 PM

Quote:
the nforce has lower latency than the kt266a


Please check the latency thread in the memory section, the only person so far to post an nForce score has had higher latency than the 2 KT266a scores posted.

<font color=blue>If you don't buy Windows, then the terrorists have already won!</font color=blue> - Microsoft
April 5, 2002 9:06:41 PM

Quote:
Please check the latency thread in the memory section, the only person so far to post an nForce score has had higher latency than the 2 KT266a scores posted.


I dont think that that benchmark is right, And secondly, I didnt see a single nforce board tested, please tell me who had it for comparison.
:wink: The Cash Left In My Pocket,The BEST Benchmark :wink:
April 5, 2002 9:56:51 PM

Sorry, I may be thinking of another place, and yes I've seen what you discovered about that benchmark. It seemed off for RDRAM/DDR, but I figured that at least between the same general platform (Athlon and DDR) it'd stay the same.

<font color=blue>If you don't buy Windows, then the terrorists have already won!</font color=blue> - Microsoft
!