If Intel, for instance, chooses to pit a dual Itanium 2 system against the dual Opteron. Itanium 2s can have shitloads of L3 cache (like 6M, vs 1M in the Opteron), which is perhaps the most important performance feature of a server chip.
Keep in mind that server applications are a totally different beast from desktop/gaming apps/multimedia apps (things that most people here on slashdot are accustomed to). While a media application has a very high instruction throughput (say, 2 instructions retired per cycle, or more if you consider the SIMD part), server applications can be as slow as 1 instruction retired every 10 cycles. This is because they have poor cache locality, and they block on data from the main memory. In any case, for a server app you generally want as much cache as possible.
First, most code is not parallelizable. So you end up being limited by Amdahl's law. Let's say that half the code can be parallelized, while the other half cannot. Even if you have infinite speedup on the parallelizable part, you still only made the program twice as fast. Cell is for games, where such parallelism is abundant. For most workloads though, it's very difficult to extract thread-level parallelism or SIMD parallelism.
Second, it's still much cheaper for Intel & others to make faster processors than for everybody else to change the software. Keep in mind that compatibility made Intel what it is now, and furthermore that most software is crap.
Niagra is a server chip. It works well on OLTP, web-serving style workloads, because those have an inherent thread-level scalability and also miss to memory a lot. Instead of having a wide, out-of-order core that is unutilized most of the time, it's more efficient to have a bunch of simple, in-order cores that execute multiple threads.
That's good for sun, because they sell server stuff, but for other kinds of workloads
this approach is very innefficient. See the Piranha research paper, by Barroso et al.
The USB Mass storage class doesn't need additional drivers because it's a well-defined, well-supported communication interface. IOW, there's one driver in the kernel, handling
As for the hardware/software RAID hybrids - it's really just cost. A fully-hardware interface needs a dedicated I/O processor, and it's generally 3-400$+. Sure, it beats hybrid RAID, but it's also much more expensive and not really worth it for desktop stuff.
Everybody seems to be criticizing this product b/c it's expensive per GB. However...
1. it works with old DDR memory. You can put newer memory, but it's only gonna be clocked at 100MHz DDR (DDR-200), mostly because the SATA cable is the limiter. IOW, if you're doing an upgrade you can put your old DDR to good use.
2. Nobody says one should make it the "only" drive. It can only get you 8GB anyway. What you can do - asymmetric pairing with a regular hdd. For instance, you could have the journal of a data-ordered ext3 filesystem on this thing, and the regular part on the HDD. This way you can almost get the best of both worlds (cheap capacity *and* speed).
3. In a server configuration this thing can do miracles for throughput. Just put your DB on it, and life is gonna be so much better...
I was wondering if someone could point me to a publication in a respectable scientific (pop-sci not included) journal that disproves global warming, with proper computer simulations.
I'm aware of many publications that show the opposite.
My very strong impression is that they did a crawling when they started it, and just stayed with that database.
I'm unable to find stuff published in my field this year with google scholar (including 2 of my papers).
Re:Which RAID controller is good?
on
Basics of RAID
·
· Score: 1
Hmmm, I don't know the promise cards. I've been using a HighPoint PCI-X RocketRAID 1820A (~210$), and it's been working pretty well. Here's one review that recomends it for desktop: tweakers.
I'd do some more research before purchasing a RAID solution, though. What you basically want is a "smart" RAID card - one that does all the work by itself, without asking the main processor to compute stuff (like RAID 5 XORs).
raid for desktop - not really worth it
on
Basics of RAID
·
· Score: 2, Interesting
RAID makes a lot of sense in a server scenario, where uptime is crucial, and where the cost of hardware is smaller than the business lost by downtime.
As far as desktops are concerned - well, RAID and cheap just don't mix.
For instance, if you just want reliability, RAID 1 is enogh (2 drives). If you want reliability + fast writes, you need RAID 1+0, which means 4 drives (RAID 5 only gives faster reads). Furthermore, a good controller is crucial (from my experience, these generally cost upwards of 100$).
Finally, RAID does not subsume in any way a good backup system. I've seen cases where a damaged controller broke both harddrives in a RAID 1. However, for (most) desktop PCs, a good backup system does subsume RAID, since it's generally easy to just use a different computer, and get all the files from the backup.
For me, the excellent piece of software backuppc running on a cheap box (~300$) has worked like a charm. This might not look cheaper than RAID, but considering that I'm using just one box to back up 10 other machines, it's pretty good.
M$ is one of the big companies that hasn't done major layoffs in the States. In fact, they were hiring when very few others were. You can't really accuse them of outsourcing. Sure, there's tons of reasons to hate 'em, but this one ain't it.
They're hiring in China and in the States. I know someone who got a job there recently (went with the darkside... oh well). Their standards are very high though, so a person who doesn't get an interview with them might inadvertently conclude that they're not hiring.
Good point, but what differentiates a video clip from a true movie is that actually watching the video part is totally optional. People will likely buy video clips because they're cooler, but will only end up watching them 5% of the time.
In a free market, if you ignore a market segment, you should not have a legal way to prevent others from coming in and serving it.
Let's not forget that the free market is nothing but an idealized abstraction. This case is yet another example of market forces being incapable of driving the services/products in the right direction. Sure, it's generally much better when market forces alone take care of the situation, but this doesn't mean that when it can't we should do nothing and invoke the free market dogma.
The Cell processor still has two general-purpose threads. As for the SPEs - sure, you can use them in libs and stuff, but that means you'll be limited by Amdahl's law. If, let's say, half of your CPU resources were consumed by library stuff, and you achieved an infinite speedup on this half, you'd only have a program that's twice as fast.
Cell is, in fact, more difficult to program than the Xbox, because the worker threads have a different instruction set then the main threads. IBM was, at some point, promoting CELL as a general computing device. Now they kinda' stopped.
Slashdot Mirror
Keep in mind that server applications are a totally different beast from desktop/gaming apps/multimedia apps (things that most people here on slashdot are accustomed to). While a media application has a very high instruction throughput (say, 2 instructions retired per cycle, or more if you consider the SIMD part), server applications can be as slow as 1 instruction retired every 10 cycles. This is because they have poor cache locality, and they block on data from the main memory. In any case, for a server app you generally want as much cache as possible.
You can also go to an IP fund with it. The IP fund will pay for patent & lawyer but will also take a significant piece of the pie (around 50%).
First, most code is not parallelizable. So you end up being limited by Amdahl's law. Let's say that half the code can be parallelized, while the other half cannot. Even if you have infinite speedup on the parallelizable part, you still only made the program twice as fast. Cell is for games, where such parallelism is abundant. For most workloads though, it's very difficult to extract thread-level parallelism or SIMD parallelism.
Second, it's still much cheaper for Intel & others to make faster processors than for everybody else to change the software. Keep in mind that compatibility made Intel what it is now, and furthermore that most software is crap.
That's good for sun, because they sell server stuff, but for other kinds of workloads this approach is very innefficient. See the Piranha research paper, by Barroso et al.
Moral: don't just write off the current leader of the market yet.
In theory, the two should work together seamlessly. In practice, they don't.
As for the hardware/software RAID hybrids - it's really just cost. A fully-hardware interface needs a dedicated I/O processor, and it's generally 3-400$+. Sure, it beats hybrid RAID, but it's also much more expensive and not really worth it for desktop stuff.
1. it works with old DDR memory. You can put newer memory, but it's only gonna be clocked at 100MHz DDR (DDR-200), mostly because the SATA cable is the limiter. IOW, if you're doing an upgrade you can put your old DDR to good use.
2. Nobody says one should make it the "only" drive. It can only get you 8GB anyway. What you can do - asymmetric pairing with a regular hdd. For instance, you could have the journal of a data-ordered ext3 filesystem on this thing, and the regular part on the HDD. This way you can almost get the best of both worlds (cheap capacity *and* speed).
3. In a server configuration this thing can do miracles for throughput. Just put your DB on it, and life is gonna be so much better ...
I'm aware of many publications that show the opposite.
I'm unable to find stuff published in my field this year with google scholar (including 2 of my papers).
Hmmm, I don't know the promise cards. I've been using a HighPoint PCI-X RocketRAID 1820A (~210$), and it's been working pretty well. Here's one review that recomends it for desktop: tweakers. I'd do some more research before purchasing a RAID solution, though. What you basically want is a "smart" RAID card - one that does all the work by itself, without asking the main processor to compute stuff (like RAID 5 XORs).
As far as desktops are concerned - well, RAID and cheap just don't mix. For instance, if you just want reliability, RAID 1 is enogh (2 drives). If you want reliability + fast writes, you need RAID 1+0, which means 4 drives (RAID 5 only gives faster reads). Furthermore, a good controller is crucial (from my experience, these generally cost upwards of 100$).
Finally, RAID does not subsume in any way a good backup system. I've seen cases where a damaged controller broke both harddrives in a RAID 1. However, for (most) desktop PCs, a good backup system does subsume RAID, since it's generally easy to just use a different computer, and get all the files from the backup.
For me, the excellent piece of software backuppc running on a cheap box (~300$) has worked like a charm. This might not look cheaper than RAID, but considering that I'm using just one box to back up 10 other machines, it's pretty good.
RIAA ? Your logic is very similar to theirs (when they compute their losses due to piracy).
M$ is one of the big companies that hasn't done major layoffs in the States. In fact, they were hiring when very few others were. You can't really accuse them of outsourcing. Sure, there's tons of reasons to hate 'em, but this one ain't it.
dark fiber doesn't have the right to vote.
I believe that the "side" residual capacitance given by taller wires doesn't really impact power consumption, it's more of a crosstalk problem.
They're hiring in China and in the States. I know someone who got a job there recently (went with the darkside ... oh well). Their standards are very high though, so a person who doesn't get an interview with them might inadvertently conclude that they're not hiring.
Do you really need all these for a server/atm configuration ? Obviously, no.
Where C is the capacitance (residual capacitance), V is the voltage and F the switching frequency.
If you don't believe me, check some real research in this area, not your kindergarten texbook, like Wattch
Shorter wires do make C a little bit smaller, but the dominant part there is the gate capacitance.
Good point, but what differentiates a video clip from a true movie is that actually watching the video part is totally optional. People will likely buy video clips because they're cooler, but will only end up watching them 5% of the time.
It's all about music videos, not movies. That makes a huge difference, IMNSHO.
Let's not forget that the free market is nothing but an idealized abstraction. This case is yet another example of market forces being incapable of driving the services/products in the right direction. Sure, it's generally much better when market forces alone take care of the situation, but this doesn't mean that when it can't we should do nothing and invoke the free market dogma.
They allow xxx-rated movies on the PSP.
Cell is, in fact, more difficult to program than the Xbox, because the worker threads have a different instruction set then the main threads. IBM was, at some point, promoting CELL as a general computing device. Now they kinda' stopped.
If I understand correctly, you're referring to the off-chip L3 cache (which is considerably slower than the on-chip Itanium one)