Benefits of Large CPU Cache

[joohang]

For high-end workstations and servers, what are some of the benefits of using CPUs with unusually large cache? (i.e. 2MB, 4MB L2 cache Xeons)

Could anyone point me out to any specific situations where large cache would be beneficial and explain to me why?

For web and DBMS servers, would it make more sense to build:

1) Two servers with dual P3 800 on a ServerWorks board.
2) One server with dual P3 Xeon 800 4MB L2 Cache.

and how about:
A) Four servers with dual P3 800 on a ServerWorks board.
B) One server with quad P3 Xeon 800 4MB L2 Cache.

I'm not familiar with Xeon CPUs, so please ignore if such CPU does not exist.

Assume that other configs are the same, except for PC133 SDRAM for the dual Coppermine and RDRAM for the dual Xeon.

 

[Elledan]

A CPU cache is actually RAM which is integrated into the CPU, which makes it much faster than RAM which is placed on the motherboard. Of course, a large cache increases the price of the CPU, which is why the 'consumer'-cpu's don't have those 1+ MB caches. RAM on the motherboard is simply cheaper.

If you want to run a server which relies on quick processing, then you'll want the XEON solution, although it'll probably cost you an arm and a leg. XEON processors are pretty expensive compared to P3 CPU's, which is mostly due to the larger cache size.

 

[joohang]

Are there any good reviews/benchmarks that show the benefits of large L2 cache?

 

[Rellik]

There is one instance where then amount of cache is even more important then its speed. I read in a review about the SETI@Home client that they tested it on different cpu´s and the ones that did
best where the one´s with a cache of 1MB. Larger caches did not improve performance since the program only occupies roughly 1MB.
The best cpu performance/cost wise was the now ancient 600Mhz P3 Katmai with 133FSB(Katmai-B) It features 512KB half speed cache with a 133FSB. 2 things important for SETI. It was faster then a 800EB P3
with 256 fullspeed cache...

But it all depends on the kind of server you run. Look at Anands Server articles....

 

[joohang]

Thanks for an interesting post.

BTW, I don't recall reading Anand's article on large cache Xeons.

 

[MadRat]

Xeons have a higher internal bandwidth than P!!!'s, too. They simplified the P!!! design in comparison to the Xeon. For some reason the P!!! Xeon has a bad reputation as being a plain jane P!!!. :Q

 

[Adul]

Large caches come into play in situations such has database servers. A large cache is much more desirable then all out speed. That is why the p3 xeon line speeds are not has high has the normal p3 line. Other situations such has large caculations, large datasets used in caculation bennfit from larger cpu caches.

 

[CTho9305]

basically what adul said. If you have a 2 meg database, running a search through it will be MUCH faster if it is in the L2 cache which runs at the internal clock (well, on modern cpus) than the main memory which is 7-10 times slower

 

[Adul]

Here uis an article on L2 Cache. From dell of all places :Q

 

[jaydee]

The three Anandtech server upgrades have a little bit in them about the functionality of what large cache does as opposed to higher clockspeed and dual processor.
Server upgrade 1
Server upgrade 2
Server upgrade 3
I wish he went a little more in-depth, cause I'm still kinda fuzzy, but its a start. Anand of all people should have a pretty good understanding considering what he has to think about in terms of these upgrades, so if he's reading this, we're all waiting for you chime in Anand

edit
Wow, that Dell link really does a number on L2 cache, thanks for the link Adul.

 

[Adul]

NO problem Jaydee

 

[pm]

From the Dell article:
"In general, the faster the cache, the more chip "real estate" it uses."

Erm. I don't think so. In general the larger the cache size, the more chip real estate it uses. And, in general, the larger the cache, the slower it is.

Aside from this minor quibble that just jumped out at me, this is very interesting stuff. Thanks for a great link, Adul.

 

[Mday]

it's a matter of having data right then and there for the cpu to perform operations on.

imagine yourself using a calculator.

what is faster? you having to look back and forth for numbers, or you knowing what the numbers are and just entering them.

--

someone mention how much cache alpha cpus have

--

generally the larger the cache, the better for programs that need it. imagine if you have enough cache to cache WINDOWS or LINUX, that'd be =O... the reason cache is so little in comparison to system mem is the cost of sram which. i am keeping my eyes peeled for mram (magnetic) which is supposed to be cheaper and as fast if not faster than current SRAM implementations.

---------

generally, i would recommend you to have more than one server to handle the load in case one of them fails.

as for the numbers you provided, it depends on what the usage is going to be. let's assume everything else is the same, RAM, HDDs, etc. (taking RDR vs SD into consideration complicates the scenario a lot)...

let's consider database:

if it's a small set of people accessing the same information, the singular servers would do for both comparisons.
if it's a small set of people accessing different information, it's a toss, depending on cost.
if it's a large set of people accessing the same information, it's a toss. some would argue that if they were located in different parts of the country, different servers would be good. if they are located in the same building, the singular server would be good.
if it's a large set of people accessing different information, the multiple servers would be good.

of course servers are more or less specialized. no one should keep an applications server the same machine as a database server or a file server.

--

now let's consider a graphics computer that renders cgi for movies. it's better to have... the individual server if the software written was meticulously created to take COMPLETE AND UTTER advantage of the computer... other wise you can have individual machines rendering different parts of the movie at the same time.

 

[joohang]

That was a great read, Adul. Thank you.

So 512KB -> 2MB L2 cache results in 15% performance gain (at least according to Dell). So I guess that it would be more effective to scale out than scaling up the cache?

 

[BFG10K]

Keep in mind that when you increase cache sizes the latency usually increases as well. For most things small fast caches are better than large slow caches.