SUMMARY: Quick Question: Do T1000s or T2000 scale better than X64?

From: JayJay Florendo <>
Date: Thu Jun 05 2008 - 03:06:45 EDT
Hi everyone, sorry for the late summary.  Thanks for all who replied.

Tobias Nutt
Anthony D'Atri
Tamer Embaby
Robert Goud
Chris Hoogendyk
Dean Ross-Smith
G Raeme
Geoff Reed
Joe Fletcher

My team spent a good 2 months planning and detailing purchase of a few
servers, only to find out at the 11th hour that due to a miscommunication
with groups outside of our team, we had gone a bit over budget, and quite
possibly used up more than the available space and power.  My plan was to
attempt to reduce our footprint by swapping x86 Opterons with Niagra based
servers at a 4:1 ratio.  To determine if this was a 'sane' ratio, I read up
on the architectures, and asked the list for feedback on X86 vs. T1/T2
Sparc.  We had about 4 days to get everything together.  We came up with 8
different scenarios/hardware combinations.  Our last proposal replaced ~4
X4200s/X4600s with 1 T5140.  We had to re-engineer the entire proposed
system (racks, power, network, application licenses, installation services,

In the end, we got 'the brass' to sign off on the idea.  We reduced our
server footprint by 75% and saved about 20% on the bottom line.  We expect
more savings to come from the recurring power/rack costs in the long run.

Here are some of the Pros/Cons going with the T1K/T2K/T5K strategy based on
the discussions with the other admins who replied:

1. Pro: T2Ks can replace V210s at a 12:1 ratio.  The admin reported that
even at full load, the T2K cluster still out performed the original cluster.

2. Con: T1Ks have embedded disks not readily accessible.

3. Con: T1K/T2Ks have only 1 CPU (cannot laterally expand) -- each CPU has
only 1 FPU, so in order to improve f(x) compute power, additional CPUs would

4. Con: Entry level for 'decent' 2 CPU machines is ~15K (X4200s only cost

5. Pro: T5140/T5240 Use less power than X4600

6. Pro: X5140s are only 1U

7. Con: T1Ks have only 1 PSU

8. Pro: Some admins preferred the ILOMs on the sparc units and OBP for

9. Pro: One admin reported running an application that finished a task in
1.5 hours on a 6 core T1K, that would take 174 hours on a single X86 thread.

10.  Pro: For Java applications, threads can share the same code stack.
Providing 1 copy of the code for 128 other threads makes more efficient use
of memory than 1 copy of the code for 4 or 8 threads.

11. Con: For Image processing, there was no perceived speedup.  The CPUs
were idle, but since the application was very computationally intensive, it
didn't perform better than 2 420Rs on 2 T6320 blades.

12. Pro: One admin reported pushing over 2M PPH using 5 T2Ks while
loadtesting apache web servers.

13. Pro: T1/T2 architectures have cryptographic accelerators (great for SSL)

14. Pro: My application was primarily Web :-)

15: Pro: T1/T2 threads interleave cache read throughs between context
switches (I haven't confirmed this) -- but according to one of the admins
who replied, when a cache miss occurs the process context switches and the
cache population occurs outside of the thread's quanta, while on X86
processors, the thread waits for the cache to populate (using up CPU time,
but not really accomplishing anything).

Other things to consider:

1. Cost of 2x4 core T5140 vs. cost of 2x6 core T5140 = ~2K

2. Cost of 2x6 core T5140 vs. cost of 2x8 core T5140 = ~15K

3. Threads per 2x6 core T5140 = 96

4. Threads per 2x8 core T5140 = 128

I don't think the extra 32 threads is worth doubling the cost of the server.

So, we went with the 2x6 core T5140s.

Best regards,

JayJay Florendo
sunmanagers mailing list
Received on Thu Jun 5 03:07:57 2008

This archive was generated by hypermail 2.1.8 : Thu Mar 03 2016 - 06:44:11 EST