I explained the gains of optimizing the PDCCH (control part) of the LTE subframe in my previous article. Let’s have a look at the data part (PDSCH) and find out various ways to improve it’s efficiency. The spectral efficiency is simply the number of bits transmitted over a frequency bandwidth in a specific time and is measured in bits/s/Hz. It is proportional to throughput as the throughput is also bits per time transmitted in a certain bandwidth. From LTE’s perspective, if the number of bits transmitted in a subframe (time) over a specific number of Resource Blocks (frequency bandwidth) is high, then it will correspond to higher throughput and higher spectral efficiency. Let’s understand the various factors impacting the spectral efficiency and ways to perform LTE throughput optimization.
– Signal To Noise & Interference Ratio:
The most basic and common factor that controls the spectral efficiency and throughput is the SINR (Signal to Noise and Interference Ratio). If the SINR of a network is bad, then that puts a limit on the throughput gain that it can achieve. So, the first thing to verify is the average SINR of the network. Let’s check some of the factors that impact SINR
This one is a basic thing. If sites are too close to each other, they will have a higher tendency to interfere with each other and will require aggressive down-tilts to limit overshooting. The distance is something that is usually fixed as LTE sites mostly use the previously deployed network. So, there is not much to do at this level other than downtilts to improve SINR and reduce overshooting.
- Electrical Tilt Ports: It is better to use antennas which assign a different RET port (electrical tilt) to LTE. That provides flexibility for optimization. If the network uses the same RET port for LTE and other RATs (3G or 2G) then any change on LTE tilt will impact the other RAT and it takes away the flexibility. So, it is a good idea to keep this in mind in the design or expansion phase.
- Pa & Pb: Another thing that can be done in case of smaller inter-site distance is to use a more balanced RS power (Reference Signal). There are two parameters in LTE Pa and Pb which define the power of the Reference Signals against the other symbols e.g. PDSCH Symbols. I will just explain with an example. If the Pa is -3 and Pb is 1, then that means that the Reference Signals will be having 3 dB higher power than the PDSCH symbols. When the inter-site distance is low, then high reference signal power can result in higher interference. If the inter-site distance is large then this configuration can be helpful as a 3dB Reference Signal boost will improve coverage as LTE coverage is controlled with RSRP and RSRP is the direct outcome of RS power. However, in case of small inter-site distance, Pb and Pa values of 0 might provide a more optimized solution as in this case, the RS power will not be boosted compared to the PDSCH symbols. Moreover, the PDCCH/PDSCH symbols in which Reference Signals are present will have a slightly higher power for 2 and 4 antenna port systems. This happens because previously with 0,-3 configuration, the RS were taking the extra available power but now with the 0,0 configuration, the extra power is used by the other channels instead of RS. So, that improves the credibility of PDSCH and can result in better throughput results. This is a big topic so I am just touching it here and will cover this in more details in the future articles.
Load & Utilization
Second factor is the load in the area or cluster. Higher the load, higher the interference to the neighbouring cells. As the load increases, the power per Resource Element increases which will result in higher aggregate power in the area increasing the RSSI. For neighbouring cell, such a power is considered interference. So, if the load increases above a threshold, it is better to add another carrier or if another carrier already exists, then it will be better to offload the congested carrier and shift the load to the uncongested carrier. This can be done using Load Balancing features or by tuning the cell reselection or mobility parameters.
Sometimes, the actual traffic volume is not that high but the utilization of the cell is still very high. This is usually caused due to low signal quality as the users with bad SINR will take a lot RBs at a lower modulation. In case, the traffic is not high but utilization is still high, it is a good idea to see the TA and CQI for the cell. If TA is pretty high and CQI is below 8 (value depends on the frequency layer) then it might be a better idea to physically optimize the area or cell. Introducing PDCCH optimization also helps in such cases as it can add to the PDSCH capacity relieving congestion to an extent.
As described in my PCI planning article, if the adjacent cells with overlapping coverages have same PCI modulo 3, then there is a probability of RS interference between them. Such an interference will reduce the overall RS SINR and demodulation capability resulting in throughput degradation. So, it should be tried to avoid PCI modulo3 conflicts wherever possible. In FDD networks, it is better to ensure that time synchronization is not enabled as that adds a randomness to the system and PCI mod3 impact is reduced significantly.
– CQI & MCS Mapping:
The next step is the CQI (Channel Quality Indicator). Once the UE measures it’s SINR, it will convert it to a CQI value so it can report to the eNB. The eNB will take this CQI and map it to a MCS (Modulation & Coding Scheme) value. A higher SINR will result in a higher CQI value and consequently, a higher MCS index. As MCS increases, the throughput usually increases so we need to ensure that we have the most optimum CQI and MCS indexes for each SINR value. In LTE, there are 16 CQI indexes and 32 MCS indexes. Usually, the CQI value of below 7 is considered bad and CQI value of around 10 is considered fair.
CQI Adjustment Algorithms
The eNB adjusts the raw CQI value shared by the UE to find an optimum CQI and this provides a higher spectral efficiency. There are basically two scenarios where this comes into play
Consider a UE-1 that measures its SINR value to be around 10 dB and based on that it calculates a CQI of 9 and sends it to the eNB. Another UE, let’s call it UE-2, measures its SINR value to be around 8 dB but based on that it sends CQI of 9 as the UEs have different chipsets from different vendors and can have a different CQI value for same SINR indexes. The eNB will have two UEs with same CQI value and if the eNB provides both of them with the same MCS (for example MCS20) then it is possible that the UE-1 might be able to work with MCS20 but the UE-2 will not be able to decode MCS20 properly at 8 dB SINR. So, to address this issue, the eNB maintains another index which is like the outer loop of BLER (Block Error Rate). Most of the vendors maintain a BLER target of 10%. Now consider the same scenario, both UEs get MCS20 and UE-1 works with a BLER value of 10% but the UE-2 had lower SINR so it will have a relatively higher BLER. Let’s say, the eNB calculates the BLER to be around 13% so the eNB will lower the MCS for the UE-2 and make it 19. If the BLER still remains above 10%, the eNB will reduce it further to ensure that the BLER target is maintained.
Similarly, if the UE sends a CQI value of 8 and eNB initiates downlink data with a MCS of 16 and it finds out that the BLER value is below 10%, it will increase MCS to 17 or 18 until the BLER target is achieved. This scenario will increase the spectral efficiency and the throughput.
So, we need to ensure that CQI adjustment or dynamic CQI assignment algorithms or outer loop control based on BLER is activated to achieve maximum gains from the channel.
Another important thing is that some vendors use low CQI values initially. For example, if the UE has just accessed the cell and it shares a CQI value of 9, the eNB will treat it as a CQI of 7 and a corresponding MCS will be allocated to it. Then after subsequent transmissions, the eNB will keep monitoring BLER and once the credibility of the UE’s CQI is ensured, the eNB will converge to the effective CQI. Some vendors keep this as a hard-coded algorithm while others provide parameters to tune this and then these parameters can be tuned to limit this behaviour resulting in faster convergence and higher throughputs especially for small packet data transfers. For instance, a UE which has a small amount of data accesses the cell and gets its data within two to three TTIs (subframes), then the eNB will not have enough CQI samples to converge quickly. The same UE will try again next time and the eNB will keep using a conservative CQI and MCS for such a UE. So, if the delta for initial CQI value is reduced, such UEs will get a less conservative CQI and MCS resulting in better data rates.
Another thing that helps is the CQI periodicity or the frequency of CQI reporting from the UE. If the UE reports CQI after a large interval, then the eNB might not have the most accurate CQI to begin with and it will take longer time to converge to the optimum MCS. Usually CQI reports are shared every 40 or 80 ms but if the UE is moving or if the channel is fluctuating then 40 or 80 ms can be considered a large interval. If we shift the CQI period to a smaller value like 20ms or 10ms, then the CQI will be more accurate and that should improve the spectral efficiency. However, the lower the interval, higher the number of CQI reports and higher the PUCCH utilization. Periodic CQI reports are sent over PUCCH in uplink so if we reduce the CQI reporting interval, that will increase the load on PUCCH. This can lead to interference on PUCCH and it can also result in RRC rejections due to PUCCH congestion. eNB needs PUCCH for CQI, HARQ & SRIs so if the PUCCH is congested, then it will have to reject new incoming access requests. This can be solved by using the following two approaches
- Adaptive or Dynamic PUCCH : This is introduced by vendors to resolve the RRC Rejections due to PUCCH overload. This allows the PUCCH to expand and it can consume more Resource Blocks if required. The down side is that the PUCCH takes the Resource Blocks from the PUSCH which can then limit the uplink throughput. However, usually the networks require higher downlink capacity so uplink can be compromised to an extent.
- Adaptive CQI Period : This is another enhancement that some vendors have. This makes the CQI reporting interval dynamic and the eNB can adjust it based on the user’s characteristics. This way, if the eNB finds a UE that has no channel fluctuation (mostly stationary), it can use longer CQI reporting interval like 80ms and eNB can reduce the interval to 10ms for a UE that has high fluctuation. This provides an optimum performance gain in CQI accuracy without impacting the PUCCH load to that extent.
There is another type of CQI reports known as Aperiodic CQIs but we will discuss that in the next episode of the throughput optimization.
Adaptive BLER Targets
Firstly, lets understand the concept of BLER. It can be divided into two categories:
- Initial BLER: When the eNB sends data to the UE and UE is unable to decode it, then it will send a HARQ NACK to the eNB. A NACK means that the eNB will have to retransmit the data and this NACK is considered IBLER or Initial Block Error.
- Residual BLER: If the UE is unable to decode the data even after retransmission, the UE will send another NACK and the eNB will have to retransmit again. However, there is a limit to these retransmissions and usually they are configurable. Commonly, these retransmissions are set to 4 and after 4 retransmissions, the eNB will not retransmit at HARQ level and consider this as a Residual Block Error.
The BLER target is maintained by the IBLER so this means that the eNB tries to maintain an IBLER of 10% for each UE. RBLER is usually very low and it is supposed to be less than 0.5%. The question may arise that why don’t we reduce the IBLER further and make it low as that should reduce retransmissions. The problem here is that lowering IBLER means that we need to lower the MCS. Even a very low MCS will not ensure a linear decrease in IBLER but it will degrade throughput excessively. So, various simulations and field trials were done to come up with an optimum target of 10% for IBLER which is followed by most of the vendors.
However, recently it has been found that BLER target of 10% works fine in fair conditions but when the radio conditions are bad or good, other BLER targets provide higher gains. For instance, if the radio conditions are bad, a BLER target of 10% keeps the MCS very conservative and increasing the BELR target, increases the MCS and it provides higher throughput gains. So, such parameters can be tuned if available to get better results.
– Mobility Strategy:
One thing that can really help in increasing the throughput is the optimum mobility strategy.
Transition to Higher CQI layer
Consider two LTE layers, for instance L800 and L1800 with same bandwidth. In this case, L800 will have a higher coverage as it is a lower frequency. So, the user count on L800 will be higher compared to L1800. However, the lower frequency layer also has higher interference since it has a bigger coverage radius. So, that will result in a lower CQI and a bad throughput. L1800 throughput will usually be better even with same bandwidth because it will have better CQI. So, the most important thing is to ensure that the layer with the better CQI gets most of the traffic. This can be done in many ways and I have jotted down a few of those.
The easiest way is to give a higher priority to L1800 and that will shift most of the UEs in L1800 coverage away from L800. This will ensure better CQI for users and thus a better throughput. Another way would be to keep them on same priority and provide a frequency offset to move the users to L1800. This is more reasonable if L1800 is also getting overloaded then the amount of load to be shifted can be tuned by varying the offsets.
I prefer load shifting by cell reselection instead of handovers. If the handover thresholds are changed or frequency priority based handovers are used, then it initiates gap periods. For UE, to move from one frequency to another frequency in connected mode, it needs to measure the target frequency. In order to measure the target frequency, the UE goes into a gap mode of 6 ms. This gap mode repeats itself after every 40 or 80ms. So, if it repeats every 40ms then that means that the UE cannot be scheduled for 6ms in every 40ms. Moreover, when the UE gets data, it needs to send a HARQ ACK/NACK after 4ms. So, it means that since the eNB knows that the UE will be in gap mode so the eNB will not schedule any data for the UE 4ms before the gap mode. That makes it 10ms in each 40ms that the UE cannot be scheduled which is around 25% of the time. So, inter-frequency handovers should be minimized as it can cause a 25% degradation in throughput. Cell reselection works in idle mode so it is a much better way to move users between the layers.
Another way is to enable load balancing between the layers and ensuring that the higher CQI layer gets more load. Load balancing usually also comes in two modes
- Connected Mode: In this case, the eNB calculates the PRBs or user count and tries to maintain target load values by performing load based handovers between the layers.
- Idle Mode: In this case, the eNB sends the frequency in the RRC Release command to the UE. eNB increases the priority of the target frequency for that UE temporarily and the UE tries to reselect to that frequency in idle mode.
Once again, I prefer idle mode based load balancing as it does not introduce the inter-frequency handovers and also gets the work done. But idle mode based load balancing will not have significant impact in case the layers have different priorities since one layer already has higher priority and idle mode based load balancing also moves users by increasing the priority. So, if the UEs are not moving to higher priority layer than that means that the layer has coverage constraints and then the idle mode based load balancing will also be unable to shift the load.
Another important factor is that many times, the low band like L800 has a bigger vertical beamwidth than the corresponding higher band. This effectively means that at the same tilt value, the L800 will have a much bigger coverage foot print than the L1800. So, before making any mobility strategy, it is important to verify the antenna patterns especially the vertical beam-width for all the layers. If the beam-width of one layer is significantly wider than the other, then ensure to put a tilt offset between the two to keep an optimum and balanced coverage.
– Scheduler Fairness:
Another important factor is the scheduler type. A scheduler can work in multiple modes
Round Robin: In this mode, the scheduler provides equal resources to all users. This is not an optimum algorithm as different users have different data requirements.
Max C/I: This mode provides significantly higher resources to users in good coverage conditions. This mode can starve the cell edge users and they will not get enough data resulting in degradation in user experience.
Proportion Fair: This scheme maintains a fairness between all users maintaining a healthy resource sharing between all user types. The basic concept of this mode is to strike a balance between users and it does that by prioritizing based on CQI and data rates. So, if the CQI is high, it will give resource to that user first but since it needs to maintain a fair data rate for all users, the cell edge users will also be scheduled. This scheme is essentially a combination of both round robin and Max C/I as it provides more resources to users with higher CQI as compared to round robin but it also provides more resources to cell edge users when compared to Max C/I. Hence, it gets the name Proportional fair.
The user throughput KPI improves with Max C/I scheduler as it provides more resources to good users resulting in higher user throughput but the cell throughput is improved with Proportional Fair algorithm as it strikes a balance between all users. So, if the user throughput KPI is to be improved then the scheduler can be tilted towards Max C/I while Proportional Fair can be used if cell throughput gain is required. The optimization at this level really needs deep understanding of the scheduler’s algorithm and it also depends if the specific vendor provides the options to play with the scheduling weights.
These are the basics to improve the spectral efficiency for a network. In the next part, I will explain the features that can be used to improve throughput along with the scenarios where they will be applicable.
In case of any queries or feedback, please drop a comment below and I would love to respond and help.