Sign in to follow this  
Followers 0
wfn

VIA KT400 - inadequate PCI throughput?

73 posts in this topic

Good evening,

Firstly here's my setup:

Gigabyte GA-7VAX (VIA KT400) M/B

AMD XP 2200+

1GB Corsair DDR433 SDRAM

ATI Radeon 9700PRO

Adaptec 29160 (BIOS 3.10.0)

Seagate Cheetah 15K.3 18GB (ST318453LW)

Seagate Cheetah 73LP (ST373405LW)

Pioneer DVD-305S

misc. cards - 3COM 3C905B-TX, Adaptec FC-4300

Windows 2000 Pro SP2 / DX8.1b

Both seagates reside on the twisted-to-flat u160 compliant cable w/ u160 lvd terminator going into the u160 port on the 29160

29160 is inserted into PCI slot #2 (irq not shared with any other slot)

While benchmarking the 15K.3 with ATTO, Adaptec SCSI Bench 32 and HDTach I get roughly 70MB/s throughput in the beginning of the drive.

Running the same benchmarks on 73LP I get around 57MB/s.

Benching both drives at the same time yields a combined throughput of about 70MB/s (36MB/s for 15K.3 and 34MB/s + some change for 73LP)

Now judging from benchmarks of other people as well as SR themselves, the 15K.3 alone should do around 75MB/s in the beginning of the drive. The burst speed (cache-to-ram) reported by HDTach and SCSI Bench32 also tops out at about 70MB/s for both the 15K.3 & 73LP.

Seems like I'm running into some sort of a PCI bottleneck, don't you think?

I remember running a bunch of drives on my P3-750 on ABIT BX6-r2 440BX board with 32bit/33mhz PCI bus connected to the same Adaptec 29160 and I could get around 103MB/s throughput easily.

So what do you make of this situation? Is this a known issue? Is there a fix? I was planning on adding some more drives to this system for video work but with this sort of performance it might be a waste of time.

Thanks.

Share this post


Link to post
Share on other sites
Benching both drives at the same time yields a combined throughput of about 70MB/s (36MB/s for 15K.3 and 34MB/s + some change for 73LP)

Yep, that's VIA. 70MB/s is actually a bit higher than the earlier VIA chipsets which topped out between 60 & 65.

Various folks will feed you a lot of crap about a "VIA PCI Latency Patch" which will cause selected benchmarks, only with selected RAID controllers, to suddenly report higher throughput than Intel chipsets, but it's BS (which can be easily shown by benching with a card that the "Latency Patch" doesn't specifically screw with, such as a gigabit ethernet NIC). What you're seeing is the speed of VIA. Get an 845PE.

KC

Share this post


Link to post
Share on other sites

well.. I am in fact running the PCI latency patch at the moment and it did not improve my throughput by a bit. I also tried running optimized profiles through WPCREdit by H.Oda and it didn't do squat either.

what board / CPU / RAM combo would you recommend? I don't really need dual, but I require good memory bandwidth along with decent PCI bandwith.

I'd love to have PCI-X 133MHz or at least a 64bit / 66mhz PCI.

Thanks!

Share this post


Link to post
Share on other sites

Best choices:

1) An E7505-chipset dual Xeon board with one processor installed (e.g. SuperMicro X5DAE, or X5DA8 if you want U320 onboard). This gives you PCI-X 133 without driving up the price too insanely, but you do need a special power supply and a case that accepts Full ATX. Also, no overclocking allowed with Xeons.

2) An i845PE-chipset single P4 board. This is the mainstream choice. You don't get PCI-X but the regular 32bit/33MHz PCI bus on the 845 is the best around, you can expect between 100-110 MBps. Best choice for a standard board (since you obviously don't need IDE RAID): Asus P4PE.

KC

Asus P4PE, Intel P4-3.06 @ 3.34GHz

1GB Corsair PC3200C2

Tekram DC-390U3W U160 SCSI

2x Seagate Cheetah X15.3 73.4GB

1x Seagate Cheetah X15-36LP 36.7GB

Yamaha F1 CD-RW

Lite-On LTD-163 DVD-ROM

TI 1394 PCI card

3Com 3C996-T Gigabit Ethernet

Creative SB Audigy 2

Visiontek GeForce4 Ti4600

Dell 2000FP 20" TFT via DVI

Share this post


Link to post
Share on other sites

thanks for the tip!

any idea if SIS based boards for AMD CPUs have the same problem with PCI as the VIA KT series boards do?

Share this post


Link to post
Share on other sites

Before I gave up and transitioned to Intel (besides, their CPUs left AMD behind a while back) I tried the SiS 735 and the nForce1. Both were better than VIA, but not in Intel class (SiS was 70-80, nForce was 80ish, maybe pushing 90, MBps). The SiS was much slower in its overall system performance than the VIA (and this was versus a KT266A, it'd be slower still vs a KT400).

The nForce had a severe PCI write (rather than read) performance issue which some people say has since been rectified by BIOS updates on certain models of nForce1 motherboard only, but I don't care anymore. Why bother screwing with a low-market-share chipset when you can keep everything in your system except the mobo & CPU and go 110MBps Intel 845PE.

KC

Share this post


Link to post
Share on other sites

I'm running a Soyo Dragon plus on an AMD XP 1700+... this is a kt266a board... running ATTO on my onboard fasttrack 100 lite I am seeing above 100MB/sec writes and even higher reads (combined score of benchin my 2 x Deskstar 75gxp RAID 0 and WD 1200BB simultaneously)

I have not run any of the VIA patches or even installed any of the VIA 4-in-1 service packs. I simply have not found them necessary on winXP (im running sp1 now)

I think it is more likely that your controller is the bottle neck and not the PCI bus.

I would make sure you have the newest mobo and controller BIOS and make sure your drives are running @ u160 mode and not 80MB mode.

Share this post


Link to post
Share on other sites
(combined score of benchin my 2 x Deskstar 75gxp RAID 0 and WD 1200BB simultaneously)

Just to clarify... How exactly are you benching both arrays simultaneously?

Peoples experience with Via chipsets vary greatly. In the majority of cases, people must resort to latency patches in order to get decent storage subsystem performance. If a card is used that isn't supported by a latency patch, PCI latency may be adjusted on a per-slot basis with many available freeware apps.

Since you are not using the 4-in-1's, I assume that you do not have the updated GART drivers installed. I am also assuming that you don't have a GeForce 4 or Radeon (which appear to require the Gart driver in order to work correctly).

Thank you for your time,

Frank Russo

Share this post


Link to post
Share on other sites
Before I gave up and transitioned to Intel (besides, their CPUs left AMD behind a while back) I tried the SiS 735 and the nForce1.  Both were better than VIA, but not in Intel class (SiS was 70-80, nForce was 80ish, maybe pushing 90, MBps).  The SiS was much slower in its overall system performance than the VIA (and this was versus a KT266A, it'd be slower still vs a KT400). 

The nForce had a severe PCI write (rather than read) performance issue which some people say has since been rectified by BIOS updates on certain models of nForce1 motherboard only, but I don't care anymore.  Why bother screwing with a low-market-share chipset when you can keep everything in your system except the mobo & CPU and go 110MBps Intel 845PE.

KC

Well, nForce is no longer a low market share chipset anymore. :) Or, more specifically, nForce2. It's quite a hot sell these days.

As far as the PCI bus on nForce1 is concerned, the write performance is indeed fixed. On my system, reads can go even higher than 90 MB/s, writes only slightly lower.

Leo

Share this post


Link to post
Share on other sites
Before I gave up and transitioned to Intel (besides, their CPUs left AMD behind a while back) I tried the SiS 735 and the nForce1.  Both were better than VIA, but not in Intel class (SiS was 70-80, nForce was 80ish, maybe pushing 90, MBps).  The SiS was much slower in its overall system performance than the VIA (and this was versus a KT266A, it'd be slower still vs a KT400).

Not that I am any big fan of VIA, but Intel isn't immune to bandwidth problems either.

Share this post


Link to post
Share on other sites
Peoples experience with Via chipsets vary greatly.  In the majority of cases, people must resort to latency patches in order to get decent storage subsystem performance.  If a card is used that isn't supported by a latency patch, PCI latency may be adjusted on a per-slot basis with many available freeware apps.

The "latency patch" is just a hack that makes specific devices appear to perform better only in specific benchmarks. It does not improve system performance. Changing latency on a per-slot basis with H.oda style tools is just going to starve the rest of the bus. 845 chipsets just work, they don't require this kind of crap with register-modifying hacker device drivers loadong on startup, etc.

Since you are not using the 4-in-1's, I assume that you do not have the updated GART drivers installed.  I am also assuming that you don't have a GeForce 4 or Radeon (which appear to require the Gart driver in order to work correctly).

A separate install of so-called "4-in-1" drivers are not required for Windows XP, they are integrated.

KC

Share this post


Link to post
Share on other sites
Not that I am any big fan of VIA, but Intel isn't immune to bandwidth problems either.

I should print up a T-shirt that says "don't believe everything you read on the Register/Inquirer".

That's a totally garbled story, based in part on a famous, ancient erratum regarding the old i850/i860 and PCI write performance from RAM being limited to 90MBps (out of 110). That's not even close to the degree of PCI bus throughput limitation exhibited by VIA.

KC

Share this post


Link to post
Share on other sites
running ATTO on my onboard fasttrack 100 lite I am seeing above 100MB/sec writes and even higher reads (combined score of benchin my 2 x Deskstar 75gxp RAID 0 and WD 1200BB simultaneously)

That's because the VIA "Latency Patch" (now integrated into the so-called "4-in-1" drivers) is designed to provide artificially higher benchmark scores with specific make/models of RAID adapters, including your Promise, to counteract bad press generated by the German TecChannel site.

KC

Share this post


Link to post
Share on other sites

Wow, this is far more misinformation in a single thread than usual. A friend was reading this and ROTFLing, and informed me of it.

Anyone bashing VIA's PCI design had better be ready to bash AMD's 760MPX PCI design as well. They're both from the same spec. I'm not in the business of defending VIA, but the last eighteen months have been an eye-opener for me, ever since I got talked into coding au-ja's VIA PCI patch in a user-friendly form at viahardware.com's request. That's right, I didn't invent the patch, and I said so quite clearly in the documentation. But since nobody reads documents, I received stacks of mail from people who thought it was mine, and I was eventually forced into working directly on patching the PCI bus.

Intel's PCI has a distinct advantage over VIA and AMD offerings, but not for long. Ever since the 440BX, Intel's PCI chipset provides "PCI bus parking", where the GNT# wire of the last active card will stay enabled until someone else actually needs the bus. This is vast improvement for a device, especially ONE device, that decides it needs to hog the bus. But, guess what? The VIA KT400 and P4X400 chipsets support "PCI bus parking" too! It's just a matter of a BIOS upgrade... or, if a particular sleazy motherboard maker won't do it, then you might just need a patch. :) My 0.20 beta build #15 enables this feature automatically. That beta build is undergoing private testing, but it should eventually be posted.

VIA does have their own "patch", and it comes in two flavors. The standard version that comes in the 4-in-1 pack will barely touch the PCI bus. It's only the special "RPP1.02" patch, downloadable from viaarena.com, that will try to tweak the PCI performance. ("RPP" means "Raid Performance Patch".)

And, as for my patch... For sixteen of the last eighteen months, my concern has been for the compatibility of specific cards on the VIA PCI bus, not for speed. So, RAID owners can go ahead and call the patch a piece of crap, because I haven't been catering to them. Even my next published patch won't tweak RAID performance, at least not by default. But if I'm left alone long enough, my patch will include a Control Panel applet that will disable VIA's performance-hurting timer safely, and individual devices can have ALMOST as good performance on a pre-KT400 as if they were on a bus-parked design. Some SCSI cards demand the bus-parked design and won't respond to this, but hey, I tried.

I'm go to retreat back to my concrete bunker. Back to the flame war!

Share this post


Link to post
Share on other sites

Oh, and if any of you happen to know of a U160 or U320 SCSI disk drive whose onboard cache can deliver data at 120+ MB/second to the bus, let me know by email at 'something AT georgebreesecom'. I already spent good money on Fujitsu drives and was disappointed. Even mode-select didn't help those turkeys.

Share this post


Link to post
Share on other sites

That's all very nice and I certainly appreciate you working on the issue.

Alas when this is all said and done all I know is my single boot drive can max out the pci bus and not even get to its full sustained transfer rate.

I've been an avid AMD supporter for almost a year but it's time to get some work done and I'm afraid I'll have to switch to Intel.

Share this post


Link to post
Share on other sites
(combined score of benchin my 2 x Deskstar 75gxp RAID 0 and WD 1200BB simultaneously)

Just to clarify... How exactly are you benching both arrays simultaneously?

Peoples experience with Via chipsets vary greatly. In the majority of cases, people must resort to latency patches in order to get decent storage subsystem performance. If a card is used that isn't supported by a latency patch, PCI latency may be adjusted on a per-slot basis with many available freeware apps.

Since you are not using the 4-in-1's, I assume that you do not have the updated GART drivers installed. I am also assuming that you don't have a GeForce 4 or Radeon (which appear to require the Gart driver in order to work correctly).

Thank you for your time,

Frank Russo

I just openned 2 copies of ATTO, set the 'total length' to 32MB, selected drive C: on one instance of ATTO, and selected drive O: on the other instance, then I hit 'Start' on both.

I can only assume this is what the author of this post did.. he really didnt give any indication of how he performed his benchmarks

running ATTO on my onboard fasttrack 100 lite I am seeing above 100MB/sec writes and even higher reads (combined score of benchin my 2 x Deskstar 75gxp RAID 0 and WD 1200BB simultaneously) 

That's because the VIA "Latency Patch" (now integrated into the so-called "4-in-1" drivers) is designed to provide artificially higher benchmark scores with specific make/models of RAID adapters, including your Promise, to counteract bad press generated by the German TecChannel site.

KC

As I stated earlier I DONT HAVE THE LATENCY PATCH INSTALLED. If you mean by your statement "A separate install of so-called "4-in-1" drivers are not required for Windows XP, they are integrated. " that winXP includes drivers for VIA's AGP, IDE controllers, NB/SB, and USB then you are probably correct... but if you mean that the drivers came from VIA as part of their 4-in-1 package you are wrong. Also, you are wrong about the latency patch being part of winXP.

KC, please just give it up. My VIA system does get above 100MB/sec PCI speeds... Whether or not this is the norm I don't know because I have not tested multiple systems.

Share this post


Link to post
Share on other sites

oh, and russofris. I have a geforce3 Ti200 and a Matrox millenium in this computer.

Share this post


Link to post
Share on other sites

blakerwry,

Would it be possible for you to use SCSI Bench32 (scbnch32.exe) to test your arrays? If you're running XP it will see your arrays as SCSI devices therefore you should be able to use that simple program provided that you have ASPI drivers running.

As a side note, I tried the same setup on an EPOX 8KHA+ KT266 motherboard and I'm seeing around 68MB/s throughput on the PCI.

Threw in the same card + drives in my ancient BH6-r2 BX440 board and I get 94MB/s without any fooling around.

Share this post


Link to post
Share on other sites
As I stated earlier I DONT HAVE THE LATENCY PATCH INSTALLED. If you mean by your statement "A separate install of so-called "4-in-1" drivers are not required for Windows XP, they are integrated. " that winXP includes drivers for VIA's AGP, IDE controllers, NB/SB, and USB then you are probably correct... but if you mean that the drivers came from VIA as part of their 4-in-1 package you are wrong. 

If you bothered to read the post above yours, you'd notice that George Breese, who did the 'black market' soundblaster latency patch before VIA came out with theirs, clearly states "The standard version [of the PCI latency patch] comes in the 4-in-1 pack".

Also, you are wrong about the latency patch being part of winXP. 

I didn't say the latency patch was part of XP. I said the 4-in-1 driver pack is not required for XP. The conversation at that time was in relation to GART drivers, etc. 99.99% of users don't need or care about the PCI latency patch so it's wrong to spread rumors that the 4-in-1 drivers are needed for XP. Don't take my words out of context.

Since you believe the 4-in-1 drivers are needed for XP, you installed them, and therefore have the PCI latency patch, thus explaining your throughput as I said before*.

KC

* although it's also entirely possible that Promise has implemented a VIA PCI latency hack in their non-inbox drivers, if you installed updated drivers from Promise. In either case, an unmodified XP install on VIA is not going to get your throughput on PCI.

Share this post


Link to post
Share on other sites
I tried the SiS 735 and the nForce1. Both were better than VIA, but not in Intel class (SiS was 70-80, nForce was 80ish, maybe pushing 90, MBps). The SiS was much slower in its overall system performance than the VIA (and this was versus a KT266A, it'd be slower still vs a KT400). 

The nForce had a severe PCI write (rather than read) performance issue which some people say has since been rectified by BIOS updates on certain models of nForce1 motherboard only, but I don't care anymore. Why bother screwing with a low-market-share chipset when you can keep everything in your system except the mobo & CPU and go 110MBps Intel 845PE. 

I tend do disagree with your findings. While admitably no fan of VIA, I found the SIS 735 to equal the pci performance of Intel and then nforce one actually exceeded it. I tested using an eight drive fc-al array consisting of 8 st336605fc drives residing on a dual channel 64 bit ( backwards compatable) FC-al controller. While admittably this array would saturate any 32/33 pci bus the only purpose was to see the maximum thouroghput each spieces would deliver.

SIS 735 97 MB/s

Intel I815 97 MB/s

Intel BX 97 MB/s

( these three were all within 1 Mb/s of each other my numbers are rounded slightly)

nvidia nforce1 101 MB/s

and sadly three iterations of VIA boards ranging from the kt133 to kt 266

67-75 MB/s

I have heard that nforce boards with there USB 2.0 chip enabled do suffer problems. Oddly enough, the usb 2.0 on the Abit nv133-r is a VIA chip. At the time my board did have its USB 2.0 disabled.

Share this post


Link to post
Share on other sites
Anyone bashing VIA's PCI design had better be ready to bash AMD's 760MPX PCI design as well. They're both from the same spec.

It was my understanding that this only applies to the 32/33 mhz bus. Can you clarify this?

Share this post


Link to post
Share on other sites

KC, read my F***ing mouth I have NEVER installed 4-in-1 drivers in winXP period.

Stop saying I have!

Since I have NEVER installed the 4-in-1 drivers I also do NOT have the latency "patch".

Got it?

Good.

Share this post


Link to post
Share on other sites

My VIA Apollo Pro 266 based board gives me 120MB/s with my Adaptec 39160 or over 110MB/s with the Compaq SMART 5300. Measured with both Winbench and ATTO. And with George's latency patch and the VIA RAID patch installed. Not sure which of the two made the most difference.

The "latency patch" is just a hack that makes specific devices appear to perform better only in specific benchmarks.

Do you really think that any of the latency patches have code in them that specifically targets my 5300 controller? I doubt anybody would have considered that someone might install a Compaq 64bit/66MHz SCSI RAID controller (designed specifically to work in Compaq servers, which mainly use ServerWorks chipsets) in to a 32bit/33MHz VIA chipset based board. These RAID controllers are worth more than a complete system of this type would often be worth.

Don't get me wrong, I'm not trying to flame you. Just wanted to show that although your experiences may have shown the latency patches to be all hype, there are also cases where they really do work.

And George, nice to have you with us. Your work is one of the default answers provided when people ask 'why is my RAID so slow?'. And believe me, they ask it! Over and over and.. so on. :)

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now
Sign in to follow this  
Followers 0