How to Choose the Right Server?

Workgroup servers have undergone a tremendous change in recent times with a
lot of new technologies hitting the market. This has made it difficult for IT
decision makers to choose the right one for their organization. The demand for
servers is increasing in India by leaps and bounds, and as a result we are
seeing that apart from the hardware majors, a lot of assemblers are also leaving
the PC space and jumping into the server arena. This has led to a major
reduction in price of branded servers.

Today we are going through a phase where processor technologies are improving
continuously and as a result we can now get a powerful server in just a couple
of Lakhs. And it is so powerful that if you had bought it ten years back, it
would have been amongst the top 50 supercomputers in the world. Such servers
take up only 2 U of a server rack. This has all been made possible because of
the availability of multi-core processors. Today, if you buy a server with a
dual Xeon, you get either the power of eight or four Xeon processors, depending
on whether you bought a quad core or a dual core server. Such a server not only
gives you commendable results in number crunching operations but also help you
in consolidating your datacenter. For example, the new age processors with
multi-core technology provide you with Virtualization support. This means you
can dedicate a single core to a given virtualized OS (which is essentially a
Server) and host 7 different server OSes on a single 2 U of a dual Quad-core
server. We say 7 instead of 8, as we assume that one core would be used for the
host server running the virtualization application.

You might wonder that if one is going to host 7 full-fledged servers on a
single quad-core server then how much memory can it support to cope with the
load of all such virtualized servers running on top of it? Well, rest assured,
these servers are not only capable enough to take a huge amount of processing
load but can also support RAM of different capacities, starting from 8 GB to 64
GB. This is enough to host any kind of application, including resource hungry
virtualization setups.

Another happening segment related to servers is storage. This segment has
essentially two concepts which are gaining importance rapidly. The first is the
penetration of Serial Attached SCSI (SAS) hard-disk. Since long SCSI was the de
facto standard for Server Direct Attached Storage (DAS). But, due to certain
limitations such as the maximum number of disks per channel and total
throughout, SAS is taking its place today. You will read more about SAS and its
features, later in this article. The second concept that is gaining importance
is the use of Host Bus Adapter (HBA) cards. These are the cards that connect the
Host (computer) to storage over iSCSI, Fiber Channel, etc. Now let's discuss all
of the above points in detail.

How much RAM?

When choosing a hardware configuration, remember that each application has a
different requirement. For instance, if you plan to use your server as a Web
Server with a lot of static content (such as a Web portal) or as a file or
networked server, then the amount of RAM doesn't really matter much, assuming
you have at least 4 GB RAM. This has become standard nowadays. We came to this
conclusion after running our set of I/O, disk and Web benchmarks, on the same
server, with 4 GB and 8 GB RAM. In case of Web requests, the performance was the
same for both configurations, while in case of file server I/O, the performance
improved by 10% with higher RAM. That's hardly an increment worth spending for
extra RAM.

What matters the most in these cases are disk and network speed. Once again,
while testing the servers, we noticed that when you stress a server's disk
access performance over the network, then the first thing to get choked is your
Gigabit Ethernet card. This results in data transfer speeds of just below 1,000
Mbps. To rectify this, we joined two Ethernet cards while running network based
benchmarks. Such an arrangement gave us double the bandwidth. We've given a
detailed report on how we ran the benchmarks in the “how we tested' section. So
here's a tip: Before thinking of upgrading your processor or RAM for I/O
intensive applications, first try teaming the free LAN cards.

Some servers came with small from HDDs (Laptop HDDs), to save on space and power consumption

Which Hard Drive?

Another important parameter for I/O intensive networked servers is hard disk
performance. This is also critical for database servers. But to get optimum
performance, you require a very high capacity hard disk with maximum rpm and
throughput, and that too in large numbers. You can actually get better
performance out of five 10,000 rpm SATA HDDs than what you can get from two
15,000 rpm SAS drives. But it depends on how much you can spend and what's more
important for you. If redundancy is not important and you have three HDDs, then
you can connect all of them in RAID 0 (Stripping mode) and get some really
extraordinary performance.

Now, if redundancy is extremely important, which is the case with most users,
you can go for RAID 5 with more hard disks (let's say 6 instead of 3) so that
you can reduce the performance degradation while writing parity information, and
get performance which is equivalent to having two disks in RAID 0.

This time we got all the servers with hot swappable SMPS, the max number was 3

Specs for compute intensive servers

The two key things that matter in this are of course, RAM and CPU power.
Plan on going for lots of RAM and a high-speed CPU if you're putting the server
to heavy duty number crunching or memory and CPU intensive applications like
virtualization. Even if you don't have a high-speed CPU, make sure that the
vendor let's you upgrade to one later. Some vendors don't offer CPU upgrades in
their server configurations. The same logic also holds valid for servers which
are used for huge number crunching, for eg, for servers that will be a node of a
cluster for doing huge data processing. One very important point to note here is
the architecture that your application has been designed for. If it's designed
for Intel architecture, then don't expect excellent performance on an AMD
platform, and vice versa.

We came to this conclusion after doing a very interesting exercise. While
deciding benchmarks to be run on all servers for testing processor performance
we first thought of using HPL (High Performance Linpack) compiled on Intel
multi-core architecture. This benchmark was provided by Intel's Server division.
When we ran this benchmark on Intel servers, it gave us 27 GFlops on Dual Core
Xeons and 53 GFlops on Quad Core Xeons. But when we ran the same test on an AMD
Dual Core, we got only 4 GFlops. This was strange as the AMD machine was, till
that point, outperforming most of the other servers. We felt there was something
wrong. As a result, we approached AMD and asked them to compile and customize
HPL for their dual core architecture. However, they were not able to provide a
customized HPL on time. So, we decided to use a processor testing benchmark
which is more real life and generic. Read more about how we stressed the
processors of various servers in 'How we tested' section.

Hot swappable hard disks are today the most common redundancy feature of servers. All the servers which we received this time had feature	More RAM slots can either be used for redundancy by distributing RAM in smallest possible chunks, or it can be used for upgradability

Service and Support

While buying a server one important aspect that should never be neglected is
the kind of support and service provided by the vendor. Different vendors have
different schemes. Though most of them provide 3 years warranty, some provide
services within a specified locality, while yet others have a deeper
penetration. For eg, if your server is installed in a remote location, a vendor,
local to a particular area, might not have a service center there and may take
time to address your problem. However, the vendor of a branded server, available
globally, would be able to address your problem by the next working day. Some
vendors tie up with local, authorized service providers to cater to the needs of
people in remote areas, where they themselves don't have any existence. There
are also vendors who provide specific services for specific needs. You need to
pay for the desired service and they will make sure that your problem is
resolved within 4 hours from the time you called them. There are other vendors
who offer you 3 year warranty with 24x7 support. But if you plan to install your
server in major cities, you might as well consider local vendors who would have
branches at least in the major cities and would be more cost effective.

Manageability

 These days servers ship with embedded sensors in chassis and
motherboard, which can help you extract information using server agents and send
them to a centralized remote console. Monitoring the server's health and
well-being remotely is a key concern for all. Some of the servers we received
had Server Management port that enables you to monitor your server over the
network. The servers also included an Intel Deployment kit that has a graphical
interface and helps you configure your server easily. Also, the Intel System
Management software allows you to monitor disk space usage and has the option to
alert the administrator when updates for firmware and BIOS are required. Some
vendors also provide their own customized Server Management CD through which you
can monitor the performance and health of the server, get reports on regular
basis and install the desired OS.

Another important component of manageability is Out of Band management also
known as LOM (lights out management) which lets you remotely manage the server
irrespective of the status of server. Using this you can only configure the
settings before the OS boots, i.e. RAID, BIOS level configuration. Two popular
ways of Out of Band management include Console redirection and BMC (Baseboard
Management Controller) management card. The console redirection is done through
the serial port with mostly a modem attached to it, you can send keyboard input
and receive text output, no GUI output is supported. BMC lets you interact with
distributed management cards using IPMI (Intelligent Platform Management
Interface). BMC now comes integrated with most of the boards else you can also
add a separate BMC card.

Reliability and Fault Tolerance

Servers are critical for any establishment and hence their reliability is
crucial. Any kind of hardware failure not only causes data loss and server
downtime but also affects the business economy. To ensure that you avoid any
kind of hardware issues, it's always preferable to have redundancy features in
the server, which take care of data protection and will continue to run the
server, even if one component fails. RAID is always considered to be a vital
solution for servers. If a server is storing high-value user data it is
preferable that they use RAID 5 configuration, which can only be achieved with
minimum of 3 hard disks.

While RAID provides protection against single disk failure, you will need
your hard disk to be hot swappable to make sure that you can replace the faulty
drive without bringing your server down. So Hot Swappable HDD enables you to
replace the faulty drives with a new one while your systems are still running.
This also enables you to avoid any unnecessary downtime.

Fans are vital for any server; they help to keep the server cool, so as to
avoid any performance bottleneck. Every server is equipped with certain number
of fans which are required to keep the server cool and performing, but if one
fan goes for a toss then their might be a possibility that your server
performance will decrease as it generates more and more heat.

So a hot swappable fan will enable you to replace the faulty fan even while
the server is running. Another key requirement is to have a hot plug and
redundant power supply which make the server fault tolerant against the failure
of an SMPS or even it can take care of power failure in case of single phase
failure or power cable failure if configured properly.

There are a few other very critical aspects of reliability that must also be
kept in mind. These include a server's resistance to extreme temperatures,
humidity, and vibrations. If you're buying servers for locations across the
country, then you must ensure that the server can withstand these environmental
conditions in those areas.

We had contacted ERTL Lab, which is a govt-owned test lab for environmental
factors. However, we were not able to pursue it much due to shortage of time.
Nevertheless, it's always good to ask the vendor for some reports on how well
can it withstand environmental factors.

Hot swappable Fans are something which we saw for the first time. Two out of nine servers we received had this feature

Expandability

With the growing need of users and organization, one of the vital decisions
while buying a server would be the expandability. A server should have enough
options for expandability should the need arise in the future. As user needs are
growing, so is the data. So storage capacity is one of the most vital
expandability options. What is the maximum number of disks the server allows you
to add?

We have seen that a maximum of 8 drive slots are available in a server, while
some offer 6 drive slots and 4 is standard across others. An SAS (Serial
Attached SCSI) can accommodate up to 1600 plus devices in one channel compared
to SCSI which can accommodate only 32. SAS uses serial communication instead of
parallel method and hence is capable of providing higher speed data transfer. So
SAS is preferred over SCSI nowadays. As far as SATA is concerned it offers
lesser expansion in terms of number of disks. An external SCSI connector enables
you to connect external SCSI devices like tape drives without the need to open
the server cabinet for connecting it.

You also need to consider how many DIMM slots are supported by the server as
they help you to add additional RAM, so if your server supports 16 GB of RAM but
there are only 4 DIMM slots available that means you need to plug a 4 GB RAM in
each of the slot, which might be an expensive affair. So more the number of DIMM
slots, better it is. Also, PCI-X (64-bit) slots help you to add additional fiber
channels or RAID cards, which you might include as the need arises. Hence you
are also required to consider the number of free PCI-X slots available. RAM
upgrade capability of a server can go up to a maximum of 64 GB, while some offer
32 GB, 16 GB upgradeability was common across most of them. Any server providing
less than 16 GB of RAM upgrade option is not worth the deal, as more and more
applications are becoming memory hungry and sooner than later you will require
higher RAM support.

The maximum number of RAM slots we this time was 16 and came with an assembled server

Now with Quad Core already available in the market, processor upgradeability
is an option that must not be neglected. Almost all vendors offer processor
upgradeability, there are some who don't offer it as a field operation, even
though processor can be upgraded on the same socket.

HBAs (Host Bus Adapters) are one of the most sought options in servers.
Though not many vendors provide it as a default option, still it can be
additionally added and it helps you to add host system to other network and
storage devices. We have mainly looked into fiber optic HBA which acts as
gateway to external SAN, though there are other types as well like ATA, SAS/SATA,
eSATA etc.

Anindya Roy, Saurangshu Kanunjna, Swapnil Arora with help from Jasmine Desai,
Jayanta Ghosh, Vijay Chauhan