SPC-1 IOPS performance per GB-NAND – chart of the month

Bar chart depicting IOPS/GB-NAND, #1 is Datacore Parallel Server with ~266 IOPS/GB-NAND,
Bar chart depicting IOPS/GB-NAND, #1 is Datacore Parallel Server with ~266 IOPS/GB-NAND,
(c) 2016 Silverton Consulting, All Rights Reserved

The above is an updated chart from last months SCI newsletter StorInt™ SPC Performance Report depicting the top 10 SPC-1 submissions IOPS™ per GB-NAND. We have been searching for a while now how to depict storage system effectiveness when using SSD or other flash storage. We have used IOPS/SSD in the past but IOPS/GB-NAND looks better.

Calculating IOPS/GB-NAND

SPC-1 does not report this metric but it can be calculated by dividing IOPS by NAND storage capacity. One can find out NAND storage capacity by looking over SPC-1 full disclosure reports (FDR), totaling up the NAND storage in the configuration in all the SSDs and flash devices. This is total NAND capacity, not Total ASU (used storage) Capacity. GB-NAND reflects just what’s indicated for SSD/flash device capacity in the configuration section. This is not necessarily the device’s physical NAND capacity when over provisioned, but at least it’s available in the FDR.

DataCore Parallel Server IOPS/GB-NAND explained

The DataCore Parallel Server generated over 5M IOPS (IO’s/second) under an SPC-1 (OLTP-like) workload. And with their 54-480GB SSDs, totaling ~25.9TB of NAND capacity, it gives them just under 200 IOPS/GB-NAND. The chart in the original report was incorrect.  There we used 36-480GB SSDs or ~17.3TB of NAND to compute IOPS/GB-NAND, which gave them just under 300 IOPS/GB-NAND in the report, which was incorrect. (The full report has been since corrected and is available for re-download for subscribers to our newsletter).

The 480GB (Samsung SM863 MZ-7KM480E)SSDs were all SATA attached. Samsung lists these SSDs as V-NAND, MLC drives, rated at 97K random Reads and 26K random writes. At over 5M IOPS, it should be running close to 100% of the SSDs rated performance. However, DataCore’s Parallel Server included 2 controllers with a total of 3TB of DRAM cache,  which was then SAS connected to 4 DELL MD1220 storage arrays, each with 512GB of DRAM cache, so their total configuration had about 5TB of DRAM in it, most of which would have been used as a IO cache.

The SPC-1 submission only used 11.8TB (Total ASU capacity) of storage. All the DRAM cache help to explain how they attained 5M IOPS. Having a multi-tiered cache like DataCore-MD1220 configuration, doesn’t insure that all the cache is effectively used but even without cache tiering logic, there might not be much of an overlap between the MD1220 and Parallel Server caches. It would be more interesting to see how busy the SSDs were during this SPC-1 run.

How random the SPC-1 workload is, is subject to much speculation in the industry. Suffice it to say it’s not 100% random, but what is. Non-random OLTP workloads would tend to favor larger caches.

SPC is coming out with a new version of their benchmark with supplementary information which may shed more light on device busyness.

All SPC-1 benchmark submissions are available at storageperformance.org.

Want more?

The August 2016 and our other SPC Performance reports have much more information on SPC-1 and SPC-2 performance. Moreover, there’s a lot more performance information, covering email and other (OLTP and throughput intensive) block storage workloads, in our SAN Storage Buying Guide, available for purchase on our website. More information on file and block protocol/interface performance is included in SCI’s SAN-NAS Buying Guidealso available from our website .

~~~~

The complete SPC performance report went out in SCI’s August 2016 Storage Intelligence e-newsletter.  A copy of the report will be posted on our SCI dispatches (posts) page over the next quarter or so (if all goes well).  However, you can get the latest storage performance analysis now and subscribe to future free SCI Storage Intelligence e-newsletters, by just using the signup form in the sidebar or you can subscribe here.

 

Latest ESRP 1K-5K mailbox DB xfers/sec/disk results – chart-of-the-month

(SCIESRP120429-001) 2012 (c) Silverton Consulting, All Rights Reserved

The above chart is from our April newsletter on Microsoft Exchange 2010 Solution Reviewed Program (ESRP) results for the 1,000 (actually 1001) to 5,000 mailbox category.  We have taken the database transfers per second, normalized them for the number of disk spindles used in the run and plotted the top 10 in the chart above.

A couple of caveats first, we chart disk-only systems in this and similar charts  on disk spindle performance. Although, it probably doesn’t matter as much at this mid-range level, for other categories SSD or Flash Cache can be used to support much higher performance on a per spindle performance measure like the above.  As such, submissions with SSDs or flash cache are strictly eliminated from these spindle level performance analysis.

Another caveat, specific to this chart is that ESRP database transaction rates are somewhat driven by Jetstress parameters (specifically simulated IO rate) used during the run.  For this mid-level category, this parameter can range from a low of 0.10 to a high of 0.60 simulated IO operations per second with a median of ~0.19.  But what I find very interesting is that in the plot above we have both the lowest rate (0.10 in #6, Dell PowerEdge R510 1.5Kmbox) and the highest (0.60 for #9, HP P2000 G3 10GbE iSCSI MSA 3.2Kmbx).  So that doesn’t seem to matter much on this per spindle metric.

That being said, I always find it interesting that the database transactions per second per disk spindle varies so widely in ESRP results.  To me this says that storage subsystem technology, firmware and other characteristics can still make a significant difference in storage performance, at least in Exchange 2010 solutions.

Often we see spindle count and storage performance as highly correlated. This is definitely not the fact for mid-range ESRP (although that’s a different chart than the one above).

Next, we see disk speed (RPM) can have a high impact on storage performance especially for OLTP type workloads that look somewhat like Exchange.  However, in the above chart the middle 4 and last one (#4-7 & 10) used 10Krpm (#4,5) or slower disks.  It’s clear that disk speed doesn’t seem to impact Exchange database transactions per second per spindle either.

Thus, I am left with my original thesis that storage subsystem design and functionality can make a big difference in storage performance, especially for ESRP in this mid-level category.  The range in the top 10 contenders spanning from ~35 (Dell PowerEdge R510) to ~110 (Dell EqualLogic PS Server) speaks volumes on this issue or a multiple of over 3X from top to bottom performance on this measure.  In fact, the overall range (not shown in the chart above spans from ~3 to ~110 which is a factor of almost 37 times from worst to best performer.

Comments?

~~~~

The full ESRP 1K-5Kmailbox performance report went out in SCI’s April newsletter.  But a copy of the full report will be posted on our dispatches page sometime next month (if all goes well). However, you can get the full SPC performance analysis now and subscribe to future free newsletters by just sending us an email or using the signup form above right.

For a more extensive discussion of current SAN or block storage performance covering SPC-1 (top 30)SPC-2 (top 30) and all three levels of ESRP (top 20) results please see SCI’s SAN Storage Buying Guide available on our website.

As always, we welcome any suggestions or comments on how to improve our analysis of ESRP results or any of our other storage performance analyses.


IBM’s 120PB storage system

Susitna Glacier, Alaska by NASA Goddard Photo and Video (cc) (from Flickr)
Susitna Glacier, Alaska by NASA Goddard Photo and Video (cc) (from Flickr)

Talk about big data, Technology Review reported this week that IBM is building a 120PB storage system for some unnamed customer.  Details are sketchy and I cannot seem to find any announcement of this on IBM.com.

Hardware

It appears that the system uses 200K disk drives to support the 120PB of storage.  The disk drives are packed in a new wider rack and are water cooled.  According to the news report the new wider drive trays hold more drives than current drive trays available on the market.

For instance, HP has a hot pluggable, 100 SFF (small form factor 2.5″) disk enclosure that sits in 3U of standard rack space.  200K SFF disks would take up about 154 full racks, not counting the interconnect switching that would be required.  Unclear whether water cooling would increase the density much but I suppose a wider tray with special cooling might get you more drives per floor tile.

There was no mention of interconnect, but today’s drives use either SAS or SATA.  SAS interconnects for 200K drives would require many separate SAS busses. With an SAS expander addressing 255 drives or other expanders, one would need at least 4 SAS busses but this would have ~64K drives per bus and would not perform well.  Something more like 64-128 drives per bus would have much better performer and each drive would need dual pathing, and if we use 100 drives per SAS string, that’s 2000 SAS drive strings or at least 4000 SAS busses (dual port access to the drives).

The report mentioned GPFS as the underlying software which supports three cluster types today:

  • Shared storage cluster – where GPFS front end nodes access shared storage across the backend. This is generally SAN storage system(s).  But the requirements for high density, it doesn’t seem likely that the 120PB storage system uses SAN storage in the backend.
  • Networked based cluster – here the GPFS front end nodes talk over a LAN to a cluster of NSD (network storage director?) servers which can have access to all or some of the storage. My guess is this is what will be used in the 120PB storage system
  • Shared Network based clusters – this looks just like a bunch of NSD servers but provides access across multiple NSD clusters.

Given the above, with ~100 drives per NSD server means another 1U extra per 100 drives or (given HP drive density) 4U per 100 drives for 1000 drives and 10 IO servers per 40U rack, (not counting switching).  At this density it takes ~200 racks for 120PB of raw storage and NSD nodes or 2000 NSD nodes.

Unclear how many GPFS front end nodes would be needed on top of this but even if it were 1 GPFS frontend node for every 5 NSD nodes, we are talking another 400 GPFS frontend nodes and at 1U per server, another 10 racks or so (not counting switching).

If my calculations are correct we are talking over 210 racks with switching thrown in to support the storage.  According to IBM’s discussion on the Storage challenges for petascale systems, it probably provides ~6TB/sec of data transfer which should be easy with 200K disks but may require even more SAS busses (maybe ~10K vs. the 2K discussed above).

Software

IBM GPFS is used behind the scenes in IBM’s commercial SONAS storage system but has been around as a cluster file system designed for HPC environments for over 15 years or more now.

Given this many disk drives something needs to be done about protecting against drive failure.  IBM has been talking about declustered RAID algorithms for their next generation HPC storage system which spreads the parity across more disks and as such, speeds up rebuild time at the cost of reducing effective capacity. There was no mention of effective capacity in the report but this would be a reasonable tradeoff.  A 200K drive storage system should have a drive failure every 10 hours, on average (assuming a 2 million hour MTBF).  Let’s hope they get drive rebuild time down much below that.

The system is expected to hold around a trillion files.  Not sure but even at 1024 bytes of metadata per file, this number of files would chew up ~1PB of metadata storage space.

GPFS provides ILM (information life cycle management, or data placement based on information attributes) using automated policies and supports external storage pools outside the GPFS cluster storage.  ILM within the GPFS cluster supports file placement across different tiers of storage.

All the discussion up to now revolved around homogeneous backend storage but it’s quite possible that multiple storage tiers could also be used.  For example, a high density but slower storage tier could be combined with a low density but faster storage tier to provide a more cost effective storage system.  Although, it’s unclear whether the application (real world modeling) could readily utilize this sort of storage architecture nor whether they would care about system cost.

Nonetheless, presumably an external storage pool would be a useful adjunct to any 120PB storage system for HPC applications.

Can it be done?

Let’s see, 400 GPFS nodes, 2000 NSD nodes, and 200K drives. Seems like the hardware would be readily doable (not sure why they needed watercooling but hopefully they obtained better drive density that way).

Luckily GPFS supports Infiniband which can support 10,000 nodes within a single subnet.  Thus an Infiniband interconnect between the GPFS and NSD nodes could easily support a 2400 node cluster.

The only real question is can a GPFS software system handle 2000 NSD nodes and 400 GPFS nodes with trillions of files over 120PB of raw storage.

As a comparison here are some recent examples of scale out NAS systems:

It would seem that a 20X multiplier times a current Isilon cluster or even a 10X multiple of a currently supported SONAS system would take some software effort to work together, but seems entirely within reason.

On the other hand, Yahoo supports a 4000-node Hadoop cluster and seems to work just fine.  So from a feasability perspective, a 2500 node GPFS-NSD node system seems just a walk in the park for Hadoop.

Of course, IBM Almaden is working on project to support Hadoop over GPFS which might not be optimum for real world modeling but would nonetheless support the node count being talked about here.

——

I wish there was some real technical information on the project out on the web but I could not find any. Much of this is informed conjecture based on current GPFS system and storage hardware capabilities. But hopefully, I haven’t traveled to far astray.

Comments?