SPARC Sees Updates For Linux 4.14

[phoronix]

Phoronix: SPARC Sees Updates For Linux 4.14

While Oracle recently laid off a ton of SPARC staff (and Solaris), not everyone was let go and it's still not instantly a dead platform. With Linux 4.14 there are some SPARC improvements for those relying upon this ex-Sun hardware...

http://www.phoronix.com/scan.php?page=news_item&px=SPARC-Linux-4.14

[L_A_G]

Continuing hardware support for SPARC even after the hardware has been killed off makes sense when you think about it... If you're going to be moving your remaining SPARC+Solaris to x86 (or ARM) Linux this gives you a "soft" upgrade path where you first migrate to the other OS and then replace the underlying hardware at whatever pace fits your budget and Quality of Service requirements.

[duby229]

Continuing hardware support for SPARC even after the hardware has been killed off makes sense when you think about it... If you're going to be moving your remaining SPARC+Solaris to x86 (or ARM) Linux this gives you a "soft" upgrade path where you first migrate to the other OS and then replace the underlying hardware at whatever pace fits your budget and Quality of Service requirements.

Well, a lot of those old SPARC machines were big tin machines. they were costly and were purchased on contract. It's not very likely most SPARC machines will be able to do that legally.

[GI_Jack]

Well, a lot of those old SPARC machines were big tin machines. they were costly and were purchased on contract. It's not very likely most SPARC machines will be able to do that legally.

No, they really weren't. Most sparc machines where workstations and mid-range machines. Big-Iron is almost still exclusively IBM, z-Series, and before it s/390.

SPARC is a now defunct "mid-range" class, along with IBM Power, DEC Alpha, and the Intel Itanium.

With the AMD64 platform, and semperon CPUs AND nudged itself into this space, and with Linux, x86/Linux became a very very cheap replacement. Eventually more powerful 64-bit Xeons all but destroyed this market.

Today, larger x86 machines are virtualized and chained together, so price per flop is still far cheaper. For all but increasinlgy niche tasks the Xeon dominates the server room.

New sparc vectorized machines look nice, can they really compete with an Xeon blade server running vmware?

[duby229]

No, they really weren't. Most sparc machines where workstations and mid-range machines. Big-Iron is almost still exclusively IBM, z-Series, and before it s/390.

SPARC is a now defunct "mid-range" class, along with IBM Power, DEC Alpha, and the Intel Itanium.

With the AMD64 platform, and semperon CPUs AND nudged itself into this space, and with Linux, x86/Linux became a very very cheap replacement. Eventually more powerful 64-bit Xeons all but destroyed this market.

Today, larger x86 machines are virtualized and chained together, so price per flop is still far cheaper. For all but increasinlgy niche tasks the Xeon dominates the server room.

New sparc vectorized machines look nice, can they really compete with an Xeon blade server running vmware?

But I don't think they were generally sold retail though. I'm confident in reasoning they were sold by contract in most cases. And most of those contracts cover the entire deployment of that product from the first day to the last day.

[cb88]

No, they really weren't. Most sparc machines where workstations and mid-range machines. Big-Iron is almost still exclusively IBM, z-Series, and before it s/390.

SPARC is a now defunct "mid-range" class, along with IBM Power, DEC Alpha, and the Intel Itanium.

With the AMD64 platform, and semperon CPUs AND nudged itself into this space, and with Linux, x86/Linux became a very very cheap replacement. Eventually more powerful 64-bit Xeons all but destroyed this market.

Today, larger x86 machines are virtualized and chained together, so price per flop is still far cheaper. For all but increasinlgy niche tasks the Xeon dominates the server room.

New sparc vectorized machines look nice, can they really compete with an Xeon blade server running vmware?

Xeon dominates the cheap server room.. head to head Sparc will stomp about anything in SQL performance.. mainly because it has hardware acceleration just for that that nobody else has. This is acutally part of what was being brought up recently in Linux... I've been watching the sparc-linux mailing list and it was discussed there. Unfortunately you have to pay through the teeth to get it... that isn't a problem for government contractors though or the government itself for that matter.

Other than that Sparc has always been very scalable... which is it's whole point anyway "Scalable PRocessor ARCitecture". It is a bit dated and crufty in places, not not nearly so much as x86... but it will still scale from something similar to a pic32 all the way up to machines that can go head to head with x86. The problem with Sparc in the past and one issue that still crops up for hobbiest that are on older hardware (Sparc T1-T3 basically) is that Sun made the same mistake AMD did when it designed bulldozer... most of the cheaper Sparc hardware you can get uses CMT.... but they have been moving away from this to a more performant implementation with higher clock rates in recent releases of the architecture.

For the workloads they are targeted for CMT is actually ideal... but the machines are sluggish to interact with for single threaded tasks, Even my 1.6Ghz Ultra 45 beats the snot out of my 8 core 1.2Ghz T2000 interactivity wise. But then again that is part of the scalablity aspect of Sparc... you can tailor it to your needs more easily than x86 and that was designed in from the start.

If you look at recent news it looks like Oracle has bascically turned over the reigns of Sparc to Fujitsu.... they have new M12 servers rolling out now based on Fujitsu processors instead of Oracle's.. you see any company can actually make Sparc processors as long as they implement a conformant implementation, Intel or AMD could even make them if they wanted to (actually I think AMD would do better to make Sparc's than ARM's... ).


What I think killed Oracle's Sparc and Solaris devisions was the fact that Oracle is a horrible parallelizing company to work for... things like Ryzen ThreadRipper would never happen there (basically it started out as a side project for fun apparently within AMD if you haven't read the news on that recently)... with AMD (for from what I hear Xilinx) at least the people at the top realize the value of all the people working for them... they aren't just a means to an end they are their most valuable asset. An AMD Xilinx combo could be pretty awesome... and AMD does need to buy or get in with an FPGA company since Intel has Altera now.

[L_A_G]

Well, a lot of those old SPARC machines were big tin machines. they were costly and were purchased on contract. It's not very likely most SPARC machines will be able to do that legally.

SPARC has never been a mainframe architecture so I don't know how you've gotten the idea SPARC machines would somehow be big iron machines. We're talking about one of the older RISC architectures that were created to compete with x86 and Motorola 68k type hardware. They've been used in desktops and blade/rackmount servers and have definitely been sold as something the customer can do major software changes to. Seeing how Oracle is offering their own RHEL-based Linux distro they'll probably be offering Linux upgrades to any customers with a support contract.

[cb88]

SPARC has never been a mainframe architecture so I don't know how you've gotten the idea SPARC machines would somehow be big iron machines. We're talking about one of the older RISC architectures that were created to compete with x86 and Motorola 68k type hardware. They've been used in desktops and blade/rackmount servers and have definitely been sold as something the customer can do major software changes to. Seeing how Oracle is offering their own RHEL-based Linux distro they'll probably be offering Linux upgrades to any customers with a support contract.

Fujitsu's incarnation of Sparc definitly leans more toward HPC and big iron (I'm not sure where this "big tin" thing you guys are talking about comes from )... Sparc hasn't seen a workstation class CPU release in some 10+ years or so. They were talking up the S7 Sonoma Sparc processor as a workstation class CPU but it was probably the last straw when it fell through. That isn't to say Fujitsu's CPUs wouldn't make excellent workstaiton CPUs... that just isn't what they are sold for.

Sparc can be a mainframe processor... or did you miss the Scalable part.

[Delgarde]

For the workloads they are targeted for CMT is actually ideal... but the machines are sluggish to interact with for single threaded tasks,

Yeah, that was our problem with them too... they did a poor job of balancing parallelism with outright speed. And for the product I work on, we need both - much of the work can scale to as many threads as you can throw it, but there's also a substantial part of it that can't. And given such constraints, we're better off with processors that can complete individual tasks quicker, even if they can't do as much in parallel. Sparc's slow-but-parallel approach is probably good for more I/O-oriented workloads, but not for anything doing a lot of math...

[GI_Jack]

Xeon dominates the cheap server room.. head to head Sparc will stomp about anything in SQL performance.. mainly because it has hardware acceleration just for that that nobody else has. This is acutally part of what was being brought up recently in Linux... I've been watching the sparc-linux mailing list and it was discussed there. Unfortunately you have to pay through the teeth to get it... that isn't a problem for government contractors though or the government itself for that matter.

Other than that Sparc has always been very scalable... which is it's whole point anyway "Scalable PRocessor ARCitecture". It is a bit dated and crufty in places, not not nearly so much as x86... but it will still scale from something similar to a pic32 all the way up to machines that can go head to head with x86. The problem with Sparc in the past and one issue that still crops up for hobbiest that are on older hardware (Sparc T1-T3 basically) is that Sun made the same mistake AMD did when it designed bulldozer... most of the cheaper Sparc hardware you can get uses CMT.... but they have been moving away from this to a more performant implementation with higher clock rates in recent releases of the architecture.

For the workloads they are targeted for CMT is actually ideal... but the machines are sluggish to interact with for single threaded tasks, Even my 1.6Ghz Ultra 45 beats the snot out of my 8 core 1.2Ghz T2000 interactivity wise. But then again that is part of the scalablity aspect of Sparc... you can tailor it to your needs more easily than x86 and that was designed in from the start.

If you look at recent news it looks like Oracle has bascically turned over the reigns of Sparc to Fujitsu.... they have new M12 servers rolling out now based on Fujitsu processors instead of Oracle's.. you see any company can actually make Sparc processors as long as they implement a conformant implementation, Intel or AMD could even make them if they wanted to (actually I think AMD would do better to make Sparc's than ARM's... ).


What I think killed Oracle's Sparc and Solaris devisions was the fact that Oracle is a horrible parallelizing company to work for... things like Ryzen ThreadRipper would never happen there (basically it started out as a side project for fun apparently within AMD if you haven't read the news on that recently)... with AMD (for from what I hear Xilinx) at least the people at the top realize the value of all the people working for them... they aren't just a means to an end they are their most valuable asset. An AMD Xilinx combo could be pretty awesome... and AMD does need to buy or get in with an FPGA company since Intel has Altera now.

SPARCs have always been fine machines, and as you've noted, they found their niche. The problem with sparc was never loosing to intel in the pure FLOP department.

It was losing in the performance per $$$ department. Just about every use case besides DBs, where oracle owns the entire stack from hardware to software, Linux on x86 XEON was far cheaper. These days, however, intel will include whatever custom instruction set you want in a xeon provided you have the $$$ to lay on the table, which facebook does.