• Please review our updated Terms and Rules here

ServerWorks and Win2K Server

EverStaR

EverStaR

Experienced Member
Joined
Jan 17, 2024
Messages
198
Thought since so many people were nudging me this way I would startup a thread and see who has some great tips to share.

For those new to the background. I have a couple of Dual Pentium III Intel Server MOBO's and have built out one system and collecting the few last tibits for the other. The suggestion was to try ServerWorks and Win2K Server. They have Dual 1.2666 Processors, and one has 4GB of ram and the other 3GB. Ones a Rev 1 and the other is a Rev 3 SAI2.

So, I have been reading up on ServerWorks, so at 100,000 foot level its a chipset at BareMetal to help manage dual processors, IO, etc. Seems easy enough to implement as I have the original CD with the drivers.

So, I have no large applications that I would consider running on my server. I have Win2K Pro R4 running on one, it runs super nice!

Tell me more and what simple things you would try out if you had such a system. If Win2K already leveraged Dual Processors without ServerWorks, what made ServerWorks such a hot deal back then?

Which version of Win2K Server, Datacenter and Advanced seem like overkill.

Thanks in advance!

E
 
I have W2K server installed on a partition mostly because the Microsoft Backup/Restore for that platform had some strange quirks that prevented its running on standard W2K. Beyond that, nah.
 
Besides having support for Mac shares (which is why I kept A Win2k server box around for ages) you also have built in support for Tape backup drives, and with 3rd party software you had support for Tape drive archives (enclosures with one of more tape drives and a bunch of tapes with automatic loading). SCSI support was pretty good in Win2k server. I just love SCSI tape autoloaders (DLT, LTO, AIT, DAT) and even have one for MO optical drives. I think there was a huge drop off on supporting Parallel printers after Win2k if that matters to you. Many printer venders went bankrupt so they never had XP drivers.
 
You know what might be an interesting project, but I have no idea how you'd manage this?

Back in the day Adobe Premier Pro(think pre-CS stuff) had a "render farm" function which allowed you to offload video post-processing over the network to one or more other systems. So you could hypothetically get your windows 98 PC set up as the workstation and then these two big beefy "servers" as the render nodes.

I think it'd be really interesting to see something like this in operation again, I doubt anyone's configured such a thing for 15+ years. Yes I know modern systems do this as well, but you don't see many retro things like this being done.
 
Lol, interesting idea, sad to say, not trying that one. :) Servers seldom ever interested me. I worked in IT for years, that kind of stuff all day. I do remember when Sprint was closing down its ISP back in the day we were offered some of their servers for our lab. My team was all excited so one guy thought he was going to haul it from the DC to our little basement DC in his pickup truck. The rack was so freaking heavy and his little pickup only came up so high on the Doc. So imagine like a 7-8 foot rack fully loaded with servers trying to be dropped into his pickup. I about broke a rib I was laughing so hard. That the most exciting server story I have.

Ho Humm, I am trying to find something worthwhile and realistic to try out after I get 2K Server Loaded. So far File Server seems like the most obvious, not that I need one, but hey, that's what CF Card OS loads are good for, try and reuse.
 
....you guys tried to move a fully-loaded rack? Yeah, I'd break a rib laughing at that :p You're supposed to un-rack 'em for transit!
 
hunterjwizzard said:
....you guys tried to move a fully-loaded rack? Yeah, I'd break a rib laughing at that :p You're supposed to un-rack 'em for transit!
Click to expand...
They have wheels on them, if they are good racks. I've moved fully loaded racks around in a data center a many times. I've even had some shipped fully loaded from one site to another. Normally really heavy things like storage would come racked halfway or 2/3 just to reduce the risk of damage in transit. I've actually still got a left over anti-static bag that I can fit over an entire 48U rack for shipping. They used to give us tip indicators too, so that we'd know if it wasn't handle well by the trucking company.

Of course I wouldn't try to move one without the right equipment. A full rack can easily weigh 1500lbs or more. I'm not going to try and drop in into a pickup for sure :)
 
Needless to say, they figured it out, they got a real truck and wheeled it on and then off on the other end. Yes, they had wheels.
 
hunterjwizzard said:
....you guys tried to move a fully-loaded rack? Yeah, I'd break a rib laughing at that :p You're supposed to un-rack 'em for transit!
Click to expand...
Yep, many times have I rolled a fully loaded rack. Biggest two were a couple of SGI Altix systems: a 20CPU Altix 3700 and a 32CPU Altix 350. 1,800 lbs for the 3700 and 2,200 for the 350.
 
Last edited:
EverStaR said:
Tell me more and what simple things you would try out if you had such a system. If Win2K already leveraged Dual Processors without ServerWorks, what made ServerWorks such a hot deal back then?
Click to expand...

The chipset is just the glue logic that handles stuff like memory I/O, access to expansion slots, onboard devices, etc. This is separate from whether or not the OS supports dual processors. Both the chipset and the OS must support dual processors in order to work properly.

ServerWorks became hot in the market at a time when Intel's chipsets were not doing so well. At the time, Intel had signed an agreement with Rambus, Inc. to use RDRAM on all their chipsets - both single CPU and multiple CPU. RDRAM was incredibly expensive, mostly due to royalties that had to be paid to Rambus, Inc. Intel tried to develop an intermediate chip called the Memory Translator Hub (MTH) that would convert the native RDRAM serial interface to the parallel interface required by SDRAM, but this chip had a major flaw in it and Intel had to do a recall of their i840 and i820 chipsets.

This left a hole in the market for someone else to move in. At the time, the only real game in town as BroadCom's ServerWorks ServerSet chipsets. These chipsets natively used SDRAM without the complexity of RDRAM or an MTH, allowing for cheaper overall system design.

But the ServerWorks chipsets aren't without flaws themselves. The earliest versions of the Pentium III chipset, like what is found on my Supermicro P3TDLE motherboard, has an issue whereby if the PCI-X bus is fully saturated, the system will lock up. This is easy enough to do when you have a 1Gb NIC and U160 or U320 SCSI installed in the PCI-X slots.

For ServerWorks, it was just a matter of being in the right place at the right time.
 
Back in 2002 you would need one hell of a SCSI RAID to max out a 133Mhz PCIX bus.

Serverworks motherboards seemed to be meant for small businesses. The good part is they could use cheap SDRAM and supported Tualatin P3's, bad being no AGP port.
 
Unknown_K said:
Back in 2002 you would need one hell of a SCSI RAID to max out a 133Mhz PCIX bus.

Serverworks motherboards seemed to be meant for small businesses. The good part is they could use cheap SDRAM and supported Tualatin P3's, bad being no AGP port.
Click to expand...
Early PCI-X chipsets, including most Pentium III class boards only supported 64bit 66MHz speeds which maxes out around 512MB/s. 1Gb ethernet consumes about 12.5MB/s, and full duplex SCSI U320 will consume 640MB/s theoretically.
 
1GB ethernet can do 125MBs not 12.5 (that's 100Mb max speed).

U320 is rated at 320Mbs max. SCSI is half duplex.

Seagate Cheetah 300Gb 3.5" 15K.6 drive can hit 150MB/s and came out in 2008. HD's in 2002 are much slower so you need a big bunch of them to saturate a U320 card.
 
ST336607LW drives could hit the mid 60 MB/s range, six of those in RAID5 could easily saturate an Ultra 320 SCSI bus.

Getting gigabit speeds out of a gigabit NIC on a Pentium III was a pipedream unless you had one of the 1 GHz+ models, or one of those really expensive NICs with an integrated DSP to offload some or all of the packet shuffling.

PCI-X was rather pointless on the PIII, the bus bandwidth was double the FSB bandwidth. Unless you were doing card to card transfers, it was mostly a waste.
 
I've dealt with fast packet I/O on commodity hardware using specialized libraries. There is no way a Pentium III system is touching anywhere near 1GBit.
I don't have old numbers at hand but to egress 1GBit of random junk with avg. frame size out of Ethernet port you need to max out a core of a relatively modern CPU.

Getting a packet out from the wire into the kernel, copying it to userspace, processing it, generating the response, copying it back to kernel, and sending it out, is what a single P3 core is going to be doing.

The "DSP" on a NIC can extract the attributes of layers of the packet and act upon them immediately, put them in separate queues so for user app the part of preprocessing is done, but what's the point of that if you have a single core? You have no means to distribute the work.

With two slow NUMA nodes and a single fast interface things get even worse. A "processor" CPU will have to peek into NUMA domain of the "IO" CPU and I'm not sure whether in P3 case it would be even better to perform a single copy from IO CPU zone to other as soon as possible.

There are a limited number of techinques for fast network. From some BIOS settings to kernel drivers, different memory allocators and different libraries, different programming patterns, but it also requires the end aplication to align to this. The cards with capabilities eg DSP come in range, from basic RSS queuing to complicated fingerprinting and layer decapsulation and features that adapters costing thousands of $ do. None of that will make any impact if applications aren't programmed accordingly.

Leave that old Pentium 3 to do something normal. Install a Linux on it and run a normal server.
 
Unknown_K said:
1GB ethernet can do 125MBs not 12.5 (that's 100Mb max speed).

U320 is rated at 320Mbs max. SCSI is half duplex.
Click to expand...

You're right. Thanks for the correction. Given those numbers, 320MB/s + 125MB/s is under the 512MB/s of the 64-bit 66MHz PCI-X bus. I guess it would be difficult to saturate that bus.
 
The bus load is 250MB/s because you're ingressing at 1GBit and egressing at 1Gbit.

In reality this platform was paired with a 100MBit interface and expected to process that kind of load inside some application.

GiGaBiTe said:
Getting gigabit speeds out of a gigabit NIC on a Pentium III was a pipedream unless you had one of the 1 GHz+ models, or one of those really expensive NICs with an integrated DSP to offload some or all of the packet shuffling.
Click to expand...

Custom traffic generator that just loads RNG junk into predefined template frame, Intel Q6600, DPDK EAL with appropriate PMD for 4x1Gbit intel NIC, so single PCI dev and 4 vdevs, using port 0 TX for gen and port 1 RX for receive test, line rate speed at near max core usage on 2 cores.
So ~ 250% of clock speed of some "fast" P3, more cache, 4 CPUs on one domain. 64-bit arch and memory hugepages.

I mean I can't say I'm not curious to see someone performing a network load test on Pentium 3 both in classic interrupt driven TCP/IP stack based application and in poll-mode "I/O is king" too.

As you said card to card might be doable because you're removing half of the hardware story and entirety of software side, but that's not a server application.
 
You must log in or register to reply here.
Back
Top