The Advent of the 48-way Server: The x3755 M3
The x86 market moves in lock-step with the evolution of x86 chip family from Intel and its spirited competitor, AMD, because the processor chip is easily the most expensive component of an x86 server. So with each generation of the chip, IBM, Dell, Hewlett-Packard and the rest announce a new top-end 4-socket server. IBM’s new offering in here is the x3755 M3. I discussed it’s virtues with Jeff Falkanger, Worldwide Senior Product Marketing Manager for System x.
Bloor: So what’s the basic specification of the x3755 M3 at a bare bones level?
Falkanger: It has a 2U chassis with up to four 12-core AMD Opteron 6000 Series processors. It can support up to 512 GB memory; 32 DIMM slots with 16 GB per slot. Also it can support up to 16 TB of local storage. There are 4 PCle slots and, optionally, it can include the VMware ESXi 4.1 embedded hypervisor.
Bloor: And which specific markets does it target?
Falkanger: It’s targeted at mid-market and large enterprise customers.
Bloor: Why has IBM favored AMD processors with this model?
Falkanger: With Intel processors, there’s a significant price premium on servers that kicks in when you go from two processors to four. What AMD has done with this architecture is to use the same processor pricing for both the two processor and four processor systems. So what we have is a four processor system which is way less expensive than any Intel-based 4 processor system. And because of this, we’ve tried to maximize the price performance value in this platform.
This will be particularly attractive to mid-market companies that were unwilling to pay that premium for the 4 processor Intel systems. We see a lot of clients trying to stay on the two processor systems entirely because of cost, when in fact a 4-socket system would give them better performance and memory scalability. So cost is an issue, and this platform eliminates that 4 processor cost premium.
Bloor: How unique is IBM’s offering?
Falkanger: Dell has a similar platform called the R815, which has also taken the AMD architecture and put it in a 2U size server. HP has not. It elected to stay with a 4U sized server that’s twice as big and more expensive. Dell has compromised to some degree on the server it’s built. So, for example, they cannot put four of the top performing CPUs into their server. Compared to Dell’s R815, we’ve got redundant hot-swap fans and they don’t. We also have hot-swap hard drives with RAID 5 for reliability. And it’s competitively priced.
Dell doesn’t achieve top performance in this architecture. We beat both Dell and HP in all the major published four-socket AMD benchmarks.
Bloor: What do you see as the main application areas for this server?
Falkanger: We see three primarily areas of application:
- Database systems
- High Performance Computing (HPC)
Bloor: So with database systems, are we are talking about traditional OLTP; ERM, CRM etc. plus BI.
Bloor: So how do you manage data storage for such applications.
Falkanger: Locally we’ve got up to 16 TB of storage capability, but in most circumstances we expect data to be stored in a SAN with Fibre Channel access. We used 3.5 inch drives, because they’re low cost / high density, expecting there will be some smaller applications where customers take advantage of that and run it all on the server. Also we have support for Converged Network Adapters, which do both Fibre Channel and Ethernet, which is fairly new technology. We had it a little earlier on the blades.
There are two pieces to this puzzle. You have to have the Adapter and you also have to have Network Switches that are smart. The Switches have been the lagging component, but in this quarter there are products coming out from Cisco and Brocade that work with our Adapters. Right now you’ll have discreet ports coming out of the switch, for Fiber Channel and for Ethernet. But ultimately the switch vendors will provide pluggable ports where you can choose either.
Bloor: The second area of application you mentioned is virtualization. So roughly how many virtual machines do you expect a server like this to accommodate?
Falkanger: That’s a hard thing to measure, except generically, and we haven’t done the work yet to put specific guidelines out. But I’ve had customers tell me that they expect to put 100 virtual machines of various flavors on this server, using 16 GB DIMMs and lots of memory. We tend to be conservative and careful in suggesting what’s possible, but you can easily put a large number of virtual machines on if they’re small.
We’re positioning this as the next step up from the 2 socket server; more memory capacity, more performance, but not at an exorbitant price, to do that next level up of virtualization and consolidation.
Bloor: My impression, from talking to VMware is that just about everyone is doing virtualization now. Is that IBM’s experience?
Falkanger: I’d say that it is pretty much everybody, but not on every thing. There’s certainly been a very rapid take-up of the technology in the past few years for some workloads. Some customers are a little bit nervous about how many workloads they put on a single machine, but in the past few years the technology has improved to where you can configure for high availability, for example putting a workload on two machines with failover in case one of them fails.
Bloor: I guess that’s what’s going to happen with VMs as we go forward. If you just consider what happens if you’re running 100 VMs on a machine and the whole machine fails, you’ll want to be able to recover from that fast and without too many complications.
Falkanger: Yes. So if you have 2 large machines your better able to manage that situation that if, say you have the same workloads spread over 6 or 7 smaller machines.
Bloor: I assume there’s no reason why you couldn’t include this server in a pool of resources if you were building a private cloud?
Falkanger: No. None at all.
Bloor: So now tell me about the x3755 in respect of high performance computing. Why is this an HPC machine?
Falkanger: HPC customers are typically looking for the maximum performance per dollar. So floorspace, or if you like, density, is a concern and power is a concern. However, we see many bids that come in that simply ask for a number of cores. “What’s the dollar amount for this number of cores?” We also see bids based on “this much” aggregate floating point capacity or alternatively, they may have a specific workload they want to address.
The x3755 provides a number of things. If we look at a traditional Intel 2 socket 1U system, it will have 12 cores in it. In a 2U footprint it would have 24 cores, whereas this system has 48. These cores offer very high performance. In fact the floating point performance, which tends to be very important for HPC applications, is better on the AMD processors than on Intel’s. Our benchmarks put is at about 8.5 percent better.
So we have the combination of the 2U density (i.e the core density per rack), and the lower costs processors and the better performance. Some of the HPC bids are interesting in that they’ll ask for an aggregate performance and ask competing vendors to figure out what fits. Now that there’s not this big price difference between the low-end and the top-end, what we have is an ability to provide that aggregate performance at a much lower price.
Bloor: So I guess most customers wont be buying in units of just one or two?
Falkanger: Well, I’ve seen everything from 2 to 150. Higher numbers of units is common. The way these HPC workloads are deployed is different than your normal enterprise deployment. The networking part of the solution is very important and often a big part of the cost of these solutions is the networking.
In some of the bids we’ve been able to reduce the number of network adapters, which also impacts the number of network switches you need, and this is simply because we can use a 4-socket server rather than two 2-socket servers. It’s not every case, but certainly we’ve had a some where we’ve been able to cut the number of 10 gigabit Ethernet adapters in half based on the customer’s traffic requirements.
Bloor: Thank you for your time.