Upgrading RAM: ECC or non ECC?

[Handruin]

I have a Dell PowerEdge sc420 (cheapo server) that I'm making use of for file serving, real time media recoding (Tversity), and other basic file backup means. It came with 512MB (2x 256MB) of ECC DDR2 533MHZ and has 4 total slots with a max config of 4GB (4x1GB).

I began looking at memory prices and Crucial has some reasonable prices for 2GB kits (2x1GB) in both ECC and non ECC flavors that are claimed to be compatible with the system. The price difference is small ($7) between them, with ECC being the more expensive.

Every where I've read says that the server accepts either type of RAM, just not mixed. Is the extra $7 worth ECC RAM for otherwise the same specs (or $14 since I'll be buying 4GB)? I'm mainly curious if people think ECC RAM is worth any reasonable price difference and also keeping in perspective that I paid $220 for this sc420 back in 2005.

$26
non ECC:
# Module Size: 2GB kit (1GBx2)
# Package: 240-pin DIMM
# Feature: DDR2 PC2-5300
# Specs: DDR2 PC2-5300 • CL=5 • Unbuffered • NON-ECC • DDR2-667 • 1.8V • 128Meg x 64

$33
ECC
# Module Size: 2GB kit (1GBx2)
# Package: 240-pin DIMM
# Feature: DDR2 PC2-5300
# Specs: DDR2 PC2-5300 • CL=5 • Unbuffered • ECC • DDR2-667 • 1.8V • 128Meg x 72 •

 

[ddrueding]

The only systems I've run ECC memory on are those that require it. I've never had system instability/file system corruption/etc that could be attributed to memory issues. I don't think ECC is worth much of anything.

 

[Handruin]

That's kind of what I thought also. I've never had any problems with this system in the few years it's been running, but it is of course using ECC memory so I thought maybe I should stick with it. It's not like $14 is going to break the bank, but if it isn't worth much like you say, then why bother.

 

[P5-133XL]

It really depends on how reliable the system needs to be. For anything in the normal range, I would say the additional cost and the additional slow-down (ECC RAM is always 15-30% slower than non-ECC at the same speed) is not worth it. However, if you really need the additional relibility, then that answers your question for you.

 

[Handruin]

I was reading up on ECC vs non ECC and most of the articles I found had said ECC RAM was only about 2-3% slower than non ECC. Where have you seen 15-30% slower? I didn't realize it was that much slower.

Also, is there merit to using a 64-bit OS with just 4GB of RAM (just in the scope of making the most use of RAM, not any other complications or possible issues with a 64 bit OS)?

 

[P5-133XL]

I got my figures from a seminar by IBM promoting chipkill around 10 years ago. They were saying that regular ECC has a 15%-30% penalty while chipkill was producing a 3%-5% penalty. Doing a google seach, I could not confirm those numbers

 

[blakerwry]

I've never heard of ECC memory being slower. The data path is wider to account for error checking/correction (72 vs 64 bits). So, running at the same frequency and timings should result in the same speeds.

Now, what I have seen being slower is registered memory (which is often associated with ECC, but not the same), but this should be clearly stated in the memory timings.

Just FYI, I had trouble getting an sc400/420 working with non-ecc memory at work. I gave up and just used ECC... perhaps I was trying to mix... dunno, but those things are picky!!! They require that modules occupy slots in order and that the memory modules are placed in size order as well...

 

[Handruin]

Thanks Blake, I ended up ordering 2GB (2x 1GB) of ECC memory for the SC420 from crucial. I had decided to keep the same type and maybe use both the existing 512MB with the new to give me the 2.5GB total. The memory only says ECC unbuffered (which I assume means it isn't registered).

 

[P5-133XL]

There is always a latency associated with ECC: Needed to actually calculate the error correcting code. There is a difference between HW ECC (i.e. chipkill) and software ECC (BIOS claculated). My point is that the slowdown is not BW associated and the real question is how much.

 

[blakerwry]

The memory only says ECC unbuffered (which I assume means it isn't registered).

Yup, registered and buffered are synonymous in the RAM world. Unbuffered (non-registered) is the typical RAM used by most PC's and one can safely assume if neither is specified that only unbuffered will work.

It is rare, in my experience, for a PC motherboard to accept buffered memory, though there are a few exceptions out there.

 

[Handruin]

I got the RAM last night and it worked out fine. It seems to have improved the performance just from the informal feel of things.

I've seen a few modern systems just yesterday that are configured with "8 x 8GB fully buffered DIMMs" for an HP blade setup. My group is trying to decide on going with a half blade populated 10U HP 7000c setup and I'm arguing towards 5 x 2U Dell 2950 IIIs for much less cost (with equal hardware configuration).

 

[Mercutio]

It is rare, in my experience, for a PC motherboard to accept buffered memory, though there are a few exceptions out there.

Some of the higher-end Intel and Asus motherboards will take buffered RAM.

 

[ddrueding]

I've seen a few modern systems just yesterday that are configured with "8 x 8GB fully buffered DIMMs" for an HP blade setup. My group is trying to decide on going with a half blade populated 10U HP 7000c setup and I'm arguing towards 5 x 2U Dell 2950 IIIs for much less cost (with equal hardware configuration).

Blades only make sense if additional rack space costs a lot of money. Even if you are renting space by the unit at a super-secure datacenter, it doesn't cost that much anymore.

 

[Handruin]

We are certainly trying to figure all that out...such as what are the advantages/disadvantages versus price for performance. These will be going in our own labs and we have space for them, but we're not an ultra-secure facility (nor do we need to be). There is an in-house cost for rack space, but I don't think it is extremely high (which includes the considerations for power and cooling).

The blades become even more complicated because of their need for an integrated SAN switches/backbone that is built into the chassis (or maybe removable cards, but you get my point...it's part of the blade setup and/or proprietary). From what I've seen in pricing, similar performance and space consumption can be had in 5 individual Dell 2950's and we'd save roughly $16,000...and that's without any discounts we might get with Dell (since we order LOTS from them). The blades offer us an environmental unknown and added complexity when compared to traditional rack iron.

 

[sechs]

I got my figures from a seminar by IBM promoting chipkill around 10 years ago. They were saying that regular ECC has a 15%-30% penalty while chipkill was producing a 3%-5% penalty. Doing a google seach, I could not confirm those numbers

I think that you have it backwards. Chipkill and similar technologies are more advanced than just run of the mill ECC, and protect against more errors (particularly, multibit errors). This makes them inherently slower. As I recall, however, IBM made special Chipkill DIMMs with an onboard processor that handled the ECC -- which would make them about as fast as regular, ECC DIMMS.

I don't think that this is a product with which anyone here will be dealing. Chipkill was born in the era of wondering how to deal with all of the errors now that we had to deal with *gigabytes* of memory.

In any case, the only real performance penalty for ECC memory is when there is an error. That's when it's probably acceptable to take the hit.

Registered memory, on the other hand, always has a performance penalty (of sorts), because reads and writes are buffered. Basically, the memory runs a clock cycle behind where it otherwsie would be.

I have ECC/registered memory in my main system, and can't tell the difference. Then again, I can't tell the difference between pedestrian and enthusiast memory timings....

 

[P5-133XL]

No, even with chipkill there was a latency cost for calculating the ECC to make sure that there was no error: It was just small (3-5%) because of using HW instead of SW/FW (15-30%) to do the calculaions. This penalty was in addition to any buffering cost.

The problem is that that was quite a long time ago, and I really have no current data. Googling, now says that the performance penalty is a relatively small 2-3% for software calculation but I really haven't been able to find any quality performance testing but just people saying it is so.

It does not surprise me that you can't tell the difference: With ergonomic testing, it has been shown that most people can't notice an improvlement lower than around 30%. So it is all in the range to benchmark testing.