You said "blah blah they cound run a 486 and do fine."
Please show me a 486 that can do EBGP, policy-based QoS (CBWFQ or LLQ) and route a couple DS3's worth of traffic. I'll just completely ignore the fact that this device, in an enterprise environment, would also need to most likely participate in some form of IGP to distribute routes into the network core.
I'm almost afraid to say it with the army of Ubuntu users on here, but I've installed fedora 10 on both an Asus Eee 901 and a Dell Mini9 for a couple different friends and it's worked fantastic.
Just don't forget, assuming solid state disk, to set noatime in fstab for the appropriate mount point(s). You'll see a massive improvement in performance, and, in theory, life span of your flash based disk.
Obviously you've never worked on a "router" outside of iptables in your dorm room.
Last week I established a single full mesh (50 sites) BGP session with my MPLS provider and CRUSHED a 3845, cpu was completely pegged. And that's just BGP, imagine a CPU intensive routing protocol like OSPF. Now throw some QoS on top of that, oh and let's enable the new Zone-based policy firewall. Should I continue?
No, the problem is cisco makes new, better, hardware and you expect them to continue to provide support and software for a piece of hardware they stopped making a decade ago. It's just not feasible.
I'd also like to point out that any REAL investment (think 4500, 6500, 12000GSR) has and/or will continue to work for 15 or 20 years.
Our elevator has a lever that does exactly that. You twist this metal pipe into some other big doo-dad and pull down on it, in the service room next to the elevator.
I just re-read that and realized how useless it was. posting anyway. GO GO KARMA!
Because they break down over time and we have to make sure they still work and will continue to work when we need them. They have components with half lifes that have to be monitored and replaced before they expire.
1.6 million processors divided by 96 cabinets gives me 16,666 physical processors per cabinet. 16,666 processors / 42 rack mount units per rack gives me 396 processors per RMU.
Either I'm missing something or they can fit 400 processors in 1U. Someone help me out here.
Your monitoring software should have told you what services were down, which should be a dead giveaway. Oh ping and smtp, that's a mail exchanger. Etc, etc.
Also, each device should have a description field to give you any additional information (vendor contact #s, responsible admin, list of countries that don't extradite in case it's that big financial database that's already gone down 3 times this week).
At that point, just create a spreadsheet with IP addresses and the services it's running. The whole point of a hostname is so you don't have to remember an IP address, not to describe every service the machine may be running.
When you have over 100 servers it really doesn't matter what you name them, you're not going to remember them all anyway. Eventually you learn the actual name is irrelevant, you just keep good spreadsheets. So have fun, name them whatever you want.
"First example: I might be writing a book in book/ and keep all images in a subdirectory book/images/. I think it is not far-fetched that I might want to work on only the images without downloading all the other, possibly huge, subdirectories."
"I've actually been quite happy with OpenID, since I have spawned far too many username/password pairs over the last 20-plus years, but it's a major chicken-and-egg problem. Hopefully someone out there will build a better mousetrap..."
Is it just me or is old timmy starting to sound like the boss from dilbert?
"And yet MD RAID1 does have better performance on single-threaded reads than single-disk reads. I think there's a bit more here than what you're describing."
That would be interesting to see, do you have any benchmarks? That doesn't jive with my (limited) understanding of how MD RAID1 handles read requests.
"Second, the same principles apply, just scaled down."
Again, a several disk raid5 solution really doesn't apply here.
"It's a cost/benefit tradeoff, and many companies do prefer to put the onus on the employee not to delete stuff that he needs"
Many companies also don't implement firewalls or don't regularly patch systems. That doesn't make it a recommended business practice. _Most_ companies take daily incremental/differential backups along with good, regular (usually weekly) full back ups. As is recommended by any competent IT professional.
"Which I explained in the previous post. You seem to be assuming that the alternative to a good backup system is no backup system? That's a reasonable misunderstanding, but I clarified that by "good" I was referring to backup frequency. Poor wording on my part."
So I should assume your other suggestion was to use RAID10 and a "bad backup solution" ? Of course the alternative to a "good backup solution" would be no backup solution. Anyway, semantics, getting nowhere.
Yes, RAID6 is fantastic, RAID-Z is also pretty amazing and a little company called lime-technology makes a great piece of software called unRAID. It lets you mix and match disks, uses a system similar to RAID-4 but each disk has it's own filesystem. So you could lose every disk but one and the data on that disk is still readable. Great for a small home setup for media and such.
"According to the Linux Software RAID FAQ, it does. More precisely, the FAQ says it performs "balanced" reads. "
Balanced read != striped read.
Balanced read means read request 1 goes to disk 1 read request 2 goes to disk 2. If read request 1 was 10x the size of read request 2 it's not very useful. Striped read means if you ask for read request 1, that single request is split between two disks. There is no comparison between the performance for single threaded reads (obviously).
"The solution was six 300 GB 15K RPM SCSI-320 drives, [...]"
I really have no idea how this has anything to do with me recommending RAID1 vs RAID0 for his HOME NAS DEVICE.
"Of course not. "
Ok well your original post said RAID10 _OR_ have a good backup system in place.
"In many environments it's a non-issue with respect to backup frequency, because the policy is that if you delete stuff you wanted to keep, you're probably screwed. "
That's insane. So if User A deletes incredibly important company information, well, it's his fault. Great, fire him. Now what?
I will however agree with you on RAID6. You take a little performance and capacity hit over RAID5 but it's worth it.
As far as RAID0 vs RAID1 read performance, it depends on whether or not the RAID controller will do striped reads on the RAID1 array. If it will perform striped RAID1 reads (which I don't believe md will), the performance will be nearly identical. Also, in practice, the controller will actually split requests between disks, so while single threaded benchmarks look bad, real world performance is actually much better.
As far as having a backup: RAID is not backup. RAID doesn't help you when the filesystem corrupts itself or you accidentally delete something. They aren't alternatives for one another. RAID is for high availability and performance and backups are for recovery.
I'm using openfiler on an hp ml570 g3 with a pair of hp dl380 g4ps running esxi 3.5 in a lab and it's been phenomenal. Using some old scavenged fiber NIC's that don't even support jumbo frames I was able to get over 500mb/s to the iscsi target on the openfiler server. It's nice to be able to build your own makeshift san out of some spare hardware from the datacenter.
Slashdot Mirror
"4 seperate copies of all the userland shared libs..."
http://blogs.vmware.com/virtualreality/2008/03/memory-overcomm.html
You said "blah blah they cound run a 486 and do fine."
Please show me a 486 that can do EBGP, policy-based QoS (CBWFQ or LLQ) and route a couple DS3's worth of traffic. I'll just completely ignore the fact that this device, in an enterprise environment, would also need to most likely participate in some form of IGP to distribute routes into the network core.
I'm almost afraid to say it with the army of Ubuntu users on here, but I've installed fedora 10 on both an Asus Eee 901 and a Dell Mini9 for a couple different friends and it's worked fantastic.
Just don't forget, assuming solid state disk, to set noatime in fstab for the appropriate mount point(s). You'll see a massive improvement in performance, and, in theory, life span of your flash based disk.
Actually it's like 7-8% smaller, they're 92% or 93% sized keyboards. Also the 901HA (or one of the 900 series) has the larger keyboard as well.
Obviously you've never worked on a "router" outside of iptables in your dorm room.
Last week I established a single full mesh (50 sites) BGP session with my MPLS provider and CRUSHED a 3845, cpu was completely pegged. And that's just BGP, imagine a CPU intensive routing protocol like OSPF. Now throw some QoS on top of that, oh and let's enable the new Zone-based policy firewall. Should I continue?
Yes, clearly, you're the expert here.
No, the problem is cisco makes new, better, hardware and you expect them to continue to provide support and software for a piece of hardware they stopped making a decade ago. It's just not feasible.
I'd also like to point out that any REAL investment (think 4500, 6500, 12000GSR) has and/or will continue to work for 15 or 20 years.
if you eat seafood and "steak" from subway, I seriously pity you.
Our elevator has a lever that does exactly that. You twist this metal pipe into some other big doo-dad and pull down on it, in the service room next to the elevator.
I just re-read that and realized how useless it was. posting anyway. GO GO KARMA!
Because they break down over time and we have to make sure they still work and will continue to work when we need them. They have components with half lifes that have to be monitored and replaced before they expire.
To put it as simply as possible: so we don't have to actually detonate the nukes to test them.
1.6 million processors divided by 96 cabinets gives me 16,666 physical processors per cabinet. 16,666 processors / 42 rack mount units per rack gives me 396 processors per RMU.
Either I'm missing something or they can fit 400 processors in 1U. Someone help me out here.
Your monitoring software should have told you what services were down, which should be a dead giveaway. Oh ping and smtp, that's a mail exchanger. Etc, etc.
Also, each device should have a description field to give you any additional information (vendor contact #s, responsible admin, list of countries that don't extradite in case it's that big financial database that's already gone down 3 times this week).
If you seriously can't bear the extra half a kilo when carrying it on your bike, you seriously need to get to the gym.
At that point, just create a spreadsheet with IP addresses and the services it's running. The whole point of a hostname is so you don't have to remember an IP address, not to describe every service the machine may be running.
When you have over 100 servers it really doesn't matter what you name them, you're not going to remember them all anyway. Eventually you learn the actual name is irrelevant, you just keep good spreadsheets. So have fun, name them whatever you want.
"First example: I might be writing a book in book/ and keep all images in a subdirectory book/images/. I think it is not far-fetched that I might want to work on only the images without downloading all the other, possibly huge, subdirectories."
check out gitweb
MOST commercial projects are a FRACTION of that size.
"I've actually been quite happy with OpenID, since I have spawned far too many username/password pairs over the last 20-plus years, but it's a major chicken-and-egg problem. Hopefully someone out there will build a better mousetrap ..."
Is it just me or is old timmy starting to sound like the boss from dilbert?
"And yet MD RAID1 does have better performance on single-threaded reads than single-disk reads. I think there's a bit more here than what you're describing."
That would be interesting to see, do you have any benchmarks? That doesn't jive with my (limited) understanding of how MD RAID1 handles read requests.
"Second, the same principles apply, just scaled down."
Again, a several disk raid5 solution really doesn't apply here.
"It's a cost/benefit tradeoff, and many companies do prefer to put the onus on the employee not to delete stuff that he needs"
Many companies also don't implement firewalls or don't regularly patch systems. That doesn't make it a recommended business practice. _Most_ companies take daily incremental/differential backups along with good, regular (usually weekly) full back ups. As is recommended by any competent IT professional.
"Which I explained in the previous post. You seem to be assuming that the alternative to a good backup system is no backup system? That's a reasonable misunderstanding, but I clarified that by "good" I was referring to backup frequency. Poor wording on my part."
So I should assume your other suggestion was to use RAID10 and a "bad backup solution" ? Of course the alternative to a "good backup solution" would be no backup solution. Anyway, semantics, getting nowhere.
Yes, RAID6 is fantastic, RAID-Z is also pretty amazing and a little company called lime-technology makes a great piece of software called unRAID. It lets you mix and match disks, uses a system similar to RAID-4 but each disk has it's own filesystem. So you could lose every disk but one and the data on that disk is still readable. Great for a small home setup for media and such.
http://lime-technology.com/
"According to the Linux Software RAID FAQ, it does. More precisely, the FAQ says it performs "balanced" reads. "
Balanced read != striped read.
Balanced read means read request 1 goes to disk 1 read request 2 goes to disk 2. If read request 1 was 10x the size of read request 2 it's not very useful. Striped read means if you ask for read request 1, that single request is split between two disks. There is no comparison between the performance for single threaded reads (obviously).
"The solution was six 300 GB 15K RPM SCSI-320 drives, [...]"
I really have no idea how this has anything to do with me recommending RAID1 vs RAID0 for his HOME NAS DEVICE.
"Of course not. "
Ok well your original post said RAID10 _OR_ have a good backup system in place.
"In many environments it's a non-issue with respect to backup frequency, because the policy is that if you delete stuff you wanted to keep, you're probably screwed. "
That's insane. So if User A deletes incredibly important company information, well, it's his fault. Great, fire him. Now what?
I will however agree with you on RAID6. You take a little performance and capacity hit over RAID5 but it's worth it.
Good lord thats amazing. I had no idea that kind of port density existed.
As far as RAID0 vs RAID1 read performance, it depends on whether or not the RAID controller will do striped reads on the RAID1 array. If it will perform striped RAID1 reads (which I don't believe md will), the performance will be nearly identical. Also, in practice, the controller will actually split requests between disks, so while single threaded benchmarks look bad, real world performance is actually much better.
As far as having a backup: RAID is not backup. RAID doesn't help you when the filesystem corrupts itself or you accidentally delete something. They aren't alternatives for one another. RAID is for high availability and performance and backups are for recovery.
I don't know, that's pretty strange. I can easily saturate fastethernet and 1GbE using it.
I'm using openfiler on an hp ml570 g3 with a pair of hp dl380 g4ps running esxi 3.5 in a lab and it's been phenomenal. Using some old scavenged fiber NIC's that don't even support jumbo frames I was able to get over 500mb/s to the iscsi target on the openfiler server. It's nice to be able to build your own makeshift san out of some spare hardware from the datacenter.
I can't believe in this day and age people would still recommend raid0.
http://www.anandtech.com/storage/showdoc.aspx?i=2101