I wish there was a more robust market for external nvme-in-box peripherals. Having to deal with vendor brain damage is never fun, and nvme devices are so small and draw little enough power that they’re really the perfect external peripheral.
Do you mean something like an external enclosure containing a backplane connecting to NVMe drives, connected to the main server over SAS 12gbps, or something else entirely?
Yeah the problem is you want more lanes than thunderbolt brings. Two nvme devices, even cheap ones, can swamp an x4 pci slot. There are various connectors designed to bring faster PCI buses out of the box, but no real standard.
There's no reason that Thunderbolt (especially optical Thunderbolt) can't be that standard. In fact, there's no reason that Thunderbolt controllers need to live in the CPU, rather than living on PCi-e HBA cards (like InfiniBand or RAID controllers do.)
It just hasn't happened yet because of a lack of enterprise demand. The pendulum hasn't swung back to all-externalizing blade servers quite yet.
Actually they currently just put a retimer chip on the backplane, attach the NVMe drives to the otherwise passive (as far as high-speed data goes) backplane, and use QSFP+ twinax cabling between the backplane and a sometimes passive "HBA" card that adapts from PCIe slot to this QSFP+ cage. Afaik they can run PCIe3 x4 per cable.
If you look for open compute hardware, you'll find various modular blade-like designs.
$1800 for a 1U server in a 24C/48T and 64GB of RAM configuration? These are incredibly competitive. How much cheaper are whitebox versions of the same thing if this is the HP premium price?
And they move dramatically less air than those little blowers they put in 1U machines. Sure at lower speeds they are quieter, but when you really need to remove the heat its hard to beat spinning at 4x the RPM, which is noisy.
Not that I've ever heard of, except for a few unaffordably high-end servers that put out absurd amounts of heat. Rackmount servers are meant for datacenters, where noise levels are generally not a concern.
Don't know a place to get it prebuilt, but building an EPYC "desktop" can be a lot like building a TR one. There are ATX server mobos. TR4 coolers should work with EPYC; Noctua explicitly says theirs do, and the physical socket and TDPs match.
You might want TR for the higher frequencies, though. I guess you could use EPYC for a very expandable 16C machine, or to pay a bit less for lower clocks, though I don't know if it ends up worth the fuss.
>The DL325 I got for $1800 is possibly below parts cost? ATM, newegg has 3960X for $1800. CPU alone.
wikipedia says the 7420P is $1250? It makes sense that the 3960X is more expensive, even though it has the same amount of cores, because it has higher clocks (3.8/4.5 base/boost vs 2.8/3.35).
Those little Dell CS24's (apparently previously facebook machines?) are cute, I own a couple, but a good 5 years ago I dropped another $200 and picked up a more recent dual 6 core machine and its about 4x faster doing everything.
I'm surprised given that he was hosting on such an old/ebay machine he didn't just drop a few hundred and get another. There is a lot of good equipment out there for cheap if you know where/what to look for.
> For background, the current server is an old ebay sourced 8x Xeon with 8GB RAM Dell. It’s actually pretty adequate, but OS upgrades require I put shoes on and walk across town to the data center.
I wonder why? Even the DELL CS24 had a remote console feature. Not saying it's not time to upgrade it of course.
Forget console. I've been upgrading OpenBSD and Linux systems remotely for ages with just the standard, runtime upgrade procedures. For example, as best I can tell my main OpenBSD system has been continuously upgraded since 2012 using the standard unattended[1] upgrade instructions, plus occasional housekeeping to remove unused system files (e.g. old libc versions). Between 2000 and 2012 there were a few changes (e.g. moving among leased and colocated hardware) requiring some reinstalls, but I otherwise followed the same pattern. I don't think I've ever maintained a single Linux instance for that long, but I've definitely upgraded across 2, possibly 3 Ubuntu LTS releases.
[1] Which is actually a manual process that requires attention, it just doesn't require console access--you're not rebooting into an installer.
I've installed Debian 4.0 beta-1 and upgraded it to until 8.x.
It had no problems but I had to switch to 64bit and there was no reliable procedure or method at that time, so I reinstalled 8. Currently the same system is running Debian 10 testing.
So it's possible to just upgrade blindly with no console.
However, a remote console is nice and useful to fix network related stuff (e.g. a switch dies or you just mess the network settings).
Note that ~5 years ago, HP started providing firmware updates and such "only to customers with a valid warranty, Care Pack Service or support agreement". They are no longer freely downloadable.
In a post ironically entitled "Customers for life", HP wrote [0]:
> ... we are in no way trying to force customers into purchasing extended coverage. That is, and always will be, a customer’s choice.
Now, this might not be an issue if all of your servers belong to $work and $work maintains active support agreements for them even after the warranty period has passed. Many of us buy servers for small businesses or even for personal use at home, however, and it would be really nice if we could keep them up-to-date with the latest firmware, security patches, and such.
At the complete opposite end of the spectrum, I can find and download any drivers, firmware, etc., for my Dell servers quickly and easily, without logging in to anything or providing any information whatsoever. If I wanted to, I could even download and install their "Repository Manager" and run my own local mirror!
WRT other vendors (unless things have changed), Cisco requires registration but not paid support (for server downloads) and IBM requires "entitlement validation". I can't speak for any other servers vendors.
EDIT: I just now checked for any updates for a Dell PowerEdge R720 (which is EoL/EoS, by the way) I have here at home, found a BIOS update from about a month ago [1], downloaded it with wget, and am now live migrating ("vMotion") the VMs running on it over to another host so that I can take the host down and install the update -- it contains microcode updates addressing several of the CVEs for Intel CPUs.
> There’s a RAID card with 2GB cache and battery included.
I can't speak for this particular card but, in the past, there were some HP RAID cards where you could not disable the RAID functionality if you wanted to.
Assume that, for whatever reason (ZFS, perhaps?), you had a bunch of drives and you just wanted to use those drives normally -- that is, you wanted the OS to see the "raw" drives, individually. Sorry, too bad!
You had two options: 1) create a bunch of one-disk RAID0 arrays (one per drive) and then let the OS use those (which isn't ideal in the ZFS case) or 2) buy a different card.
Another common thing (when it comes to ZFS) is "cross-flashing" the firmware on RAID cards. Many of the RAID cards shipped by the various server manufacturers are often just re-branded LSI cards, with or without modifications to the firmware (besides just the re-branding).
For example, you could take an IBM ServeRAID M5014 card (which is basically an LSI MegaRAID 9260-8i card with less cache) and flash it with the "original" firmware from LSI. The last I checked (which, admittedly, has been a while), it wasn't possible to do this with many of the most popular RAID cards that HP shipped.
(This probably isn't an issue if you're buying servers for $work. In my case, I buy them for $home too.)
At a previous job (at an .edu), we used HP server gear pretty extensively and, overall, I was quite happy with it. In the last 5-10 years, though, I've become "turned off" to pretty much anything from HP, though.
---
Finally, there was the recently discovered issue with HPE SAS SSDs failing catastrophically [0]:
> SSDs which were put into service at the same time will likely fail nearly simultaneously.
Nothing like having every single drive in your ENTIRE array fail at (basically) the exact same time, huh? I feel bad for the department/team that discovered this issue the hard way!
> ... results in SSD failure at 32,768 hours of operation ...
I know what you're thinking: "32,768? That's one more than the maximum value of a --"
26 comments
[ 4.1 ms ] story [ 22.4 ms ] threadIt just hasn't happened yet because of a lack of enterprise demand. The pendulum hasn't swung back to all-externalizing blade servers quite yet.
You might want TR for the higher frequencies, though. I guess you could use EPYC for a very expandable 16C machine, or to pay a bit less for lower clocks, though I don't know if it ends up worth the fuss.
Their premium comes with shit like the 200 dollar "enablement" card full of passive component.
>The DL325 I got for $1800 is possibly below parts cost? ATM, newegg has 3960X for $1800. CPU alone.
wikipedia says the 7420P is $1250? It makes sense that the 3960X is more expensive, even though it has the same amount of cores, because it has higher clocks (3.8/4.5 base/boost vs 2.8/3.35).
I'm surprised given that he was hosting on such an old/ebay machine he didn't just drop a few hundred and get another. There is a lot of good equipment out there for cheap if you know where/what to look for.
I wonder why? Even the DELL CS24 had a remote console feature. Not saying it's not time to upgrade it of course.
https://www.youtube.com/watch?v=ZtrkhIQXIp8
[1] Which is actually a manual process that requires attention, it just doesn't require console access--you're not rebooting into an installer.
It had no problems but I had to switch to 64bit and there was no reliable procedure or method at that time, so I reinstalled 8. Currently the same system is running Debian 10 testing.
So it's possible to just upgrade blindly with no console.
However, a remote console is nice and useful to fix network related stuff (e.g. a switch dies or you just mess the network settings).
In a post ironically entitled "Customers for life", HP wrote [0]:
> ... we are in no way trying to force customers into purchasing extended coverage. That is, and always will be, a customer’s choice.
Now, this might not be an issue if all of your servers belong to $work and $work maintains active support agreements for them even after the warranty period has passed. Many of us buy servers for small businesses or even for personal use at home, however, and it would be really nice if we could keep them up-to-date with the latest firmware, security patches, and such.
At the complete opposite end of the spectrum, I can find and download any drivers, firmware, etc., for my Dell servers quickly and easily, without logging in to anything or providing any information whatsoever. If I wanted to, I could even download and install their "Repository Manager" and run my own local mirror!
WRT other vendors (unless things have changed), Cisco requires registration but not paid support (for server downloads) and IBM requires "entitlement validation". I can't speak for any other servers vendors.
EDIT: I just now checked for any updates for a Dell PowerEdge R720 (which is EoL/EoS, by the way) I have here at home, found a BIOS update from about a month ago [1], downloaded it with wget, and am now live migrating ("vMotion") the VMs running on it over to another host so that I can take the host down and install the update -- it contains microcode updates addressing several of the CVEs for Intel CPUs.
[0]: https://web.archive.org/web/20140209062309/http://h30507.www...
[1]: https://www.dell.com/support/home/us/en/04/drivers/driversde...
I can't speak for this particular card but, in the past, there were some HP RAID cards where you could not disable the RAID functionality if you wanted to.
Assume that, for whatever reason (ZFS, perhaps?), you had a bunch of drives and you just wanted to use those drives normally -- that is, you wanted the OS to see the "raw" drives, individually. Sorry, too bad!
You had two options: 1) create a bunch of one-disk RAID0 arrays (one per drive) and then let the OS use those (which isn't ideal in the ZFS case) or 2) buy a different card.
Another common thing (when it comes to ZFS) is "cross-flashing" the firmware on RAID cards. Many of the RAID cards shipped by the various server manufacturers are often just re-branded LSI cards, with or without modifications to the firmware (besides just the re-branding).
For example, you could take an IBM ServeRAID M5014 card (which is basically an LSI MegaRAID 9260-8i card with less cache) and flash it with the "original" firmware from LSI. The last I checked (which, admittedly, has been a while), it wasn't possible to do this with many of the most popular RAID cards that HP shipped.
(This probably isn't an issue if you're buying servers for $work. In my case, I buy them for $home too.)
At a previous job (at an .edu), we used HP server gear pretty extensively and, overall, I was quite happy with it. In the last 5-10 years, though, I've become "turned off" to pretty much anything from HP, though.
---
Finally, there was the recently discovered issue with HPE SAS SSDs failing catastrophically [0]:
> SSDs which were put into service at the same time will likely fail nearly simultaneously.
Nothing like having every single drive in your ENTIRE array fail at (basically) the exact same time, huh? I feel bad for the department/team that discovered this issue the hard way!
> ... results in SSD failure at 32,768 hours of operation ...
I know what you're thinking: "32,768? That's one more than the maximum value of a --"
Yep.
[0]: https://news.ycombinator.com/item?id=21637516